Measuring social networks when forming information system project teams

Measuring social networks when forming information system project teams

Accepted Manuscript Measuring social networks when forming information system project teams Roberto Latorre, Javier Suarez ´ PII: DOI: Reference: S0...

3MB Sizes 2 Downloads 59 Views

Accepted Manuscript

Measuring social networks when forming information system project teams Roberto Latorre, Javier Suarez ´ PII: DOI: Reference:

S0164-1212(17)30208-X 10.1016/j.jss.2017.09.019 JSS 10041

To appear in:

The Journal of Systems & Software

Received date: Revised date: Accepted date:

7 July 2016 4 July 2017 21 September 2017

Please cite this article as: Roberto Latorre, Javier Suarez, Measuring social networks when ´ forming information system project teams, The Journal of Systems & Software (2017), doi: 10.1016/j.jss.2017.09.019

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT

Highlights • Socio-technical framework for team assignments based on social-network

CR IP T

analysis • Social networks estimating compatibility according to previous collaborations

• Social networks estimating compatibility according to social skills

• Industrial insights into the applicability and usability in real-life case-

AC

CE

PT

ED

M

AN US

studies

1

ACCEPTED MANUSCRIPT

Measuring social networks when forming information system project teams

a Escuela

CR IP T

Roberto Latorrea,∗, Javier Su´ arezb Polit´ ecnica Superior, Universidad Aut´ onoma de Madrid, 28049, Madrid, Spain b Kymatic Soluciones Inform´ aticas, Madrid, Spain

Abstract

AN US

Despite the advances in the project management field, little is known about how creating performing software teams in a systematical and repeatable way.

The technical dimension is not enough to achieve this. Software development is a complex collaborative process where people and interpersonal relationships – i.e., how people interact, behave and organize significantly – influence the project success. We define a framework to assign people to projects from this

M

socio-technical perspective. Social networks characterizing the interplay between teammates are built and analyzed to predict productive collaborations

ED

and identify adequate team-members depending on the needs of the organization and the kind of project. A noteworthy novelty of these social networks is that they estimate compatibility between coworkers according to previous

PT

collaborations, but also to individuals’ social skills. This allows analyzing the compatibility between people who have not worked together before. We present results of using the proposed framework in a multinational corporation during

CE

a more-than-two-year period. Our in-company experiments emphasize that we can significantly improve the expected outcomes characterizing and measuring

AC

the social interaction between coworkers. Social aspects discussed here may be highly relevant in the context of distributed software engineering, since it implies new challenges in the interplay among coworkers.. ∗ Corresponding

author Email address: [email protected] (Roberto Latorre)

Preprint submitted to Journal of Sytems and Software

September 22, 2017

ACCEPTED MANUSCRIPT

Keywords: Team member allocation, Team formation problem, Social skills, Social network analysis, Interpersonal relationships, Collaboration network

CR IP T

1. Introduction

Information System (IS) projects in the industrial environment are commonly considered to be successfully completed when they are delivered on time,

on budget and with the appropriate level of quality. With this definition in 5

mind, multiple factors can influence the temporal, financial and quality goals

AN US

of industrial IS projects during their lifetime (Charette, 2005). For instance,

underestimated projects commonly fail: the ones delivered with the appropriate level of quality, are often completed late and/or with over cost; while the released to meet the budget usually do not meet system requirements and are 10

unreliable systems. Another fundamental factor to complete the project successfully is an appropriate team composition (Fenton and Neil, 1999; Linberg,

M

1999; El Emam and Koru, 2008). The decision of who participates in the project has therefore a critical impact on success. Nevertheless, experience shows that, to improve the expected outcomes, the project team usually takes a backseat at the planning stage and we often focus the attention on other factors, such

ED

15

as determining the project tasks and workload, and optimizing the time-line

PT

scheduling for developing the system. Assuming the existence of such mechanisms to control these other factors, this paper addresses the team formation problem in the context of real-life industrial IS projects. Forming effective and efficient working teams is a complicated endeavor be-

CE

20

cause multiple factors influence the effective performance and productivity of a

AC

group of people working together. Traditional methods for selecting people to perform a given task are based on the candidates’ estimated performance and/or the best fitting among the individuals’ abilities and the set of abilities required

25

for the task to perform (Lievens et al., 2002). In the context of software engineering, the need of highly technical specialized skills for developing software has been subject of discussion for a long time. Thus, human resources (HR) allo-

3

ACCEPTED MANUSCRIPT

cation in software industry is mainly driven by the candidates’ know-what and know-how, i.e., technical knowledge or expertise acquired through experience 30

or formal training. For instance, these are the basis of traditional role-based

CR IP T

staffing strategies (Bandinelli et al., 1996; Zhu et al., 2006) that individually assign people to development roles with the goal of identifying the candidates with the necessary technical skills to complete each development activity with

the minimum training time and/or cost (Tsai et al., 2003; Barreto et al., 2008; 35

Otero et al., 2009; Ngo-The and Ruhe, 2009; Chen and Zhang, 2013). However, a team is more than the sum of the individual team members’ attributes. Different

AN US

individuals have different working styles and preferences, and simply bringing

together the best candidates as a function of their estimated performance or technical abilities does not ensure creating a high-performing team (Fitzpatrick 40

and Askin, 2005; Acu˜ na et al., 2008; Adams and Anantatmula, 2010; Feldt et al., 2010). Software development is an intense collaborative process (Sawyer, 2004; Luna-Reyes et al., 2005) where cooperation and social interactions among team-

M

mates (Lakhanpal, 1993; Sawyer, 2001; Gorla and Lam, 2004), or even among people belonging to different working teams (Oh et al., 2006), can significantly affect to project success. From this socio-technical perspective, an important

ED

45

step to enhance productivity and quality outcomes of IS project teams is to be able to identify compatible coworkers to improve social interactions within the

PT

team and avoid, or at least minimize, conflicts. A commonly accepted view in psychology is that the personality of an individual can be described by sets of personality traits, i.e., fixed sets of patterns

CE

50

defining the way people tend to think, feel and behave in different situations. This concept is extensively used in the HR management domain (Robertson

AC

and Smith, 2001) trying to estimate the level of development of an individual in different generic (skills or attributes of the personal behavior that can be

55

considered as a behavioral feature) and technical competencies (particular skills specifically related to a specific job or task, e.g., business analysis or project management). Software engineering research efforts in this line are directed toward identifying the required features for each development task or role (Acu˜ na 4

ACCEPTED MANUSCRIPT

and Juristo, 2004; Balthazard et al., 2004; Acu˜ na et al., 2006; Licorish and 60

MacDonell, 2015) and finding the best matching among people and these features (Drigas et al., 2004; Otero et al., 2009; Otero and Otero, 2012). Related to

CR IP T

our work, some authors claim that some competencies can explain and predict how a person interact and collaborate with other people in such a way that

variations of these behavioral and social traits can significantly influence the 65

dynamics and effective productivity of a development team (Blaskovich, 2008;

Adams and Anantatmula, 2010; Feldt et al., 2010). Under this view, it has been proposed that creating productive software development teams requires an ade-

traits (Turley and Bieman, 1995).

AN US

quate balance across team members’ technical skills, experience and personality

In spite of the increasing interest on non-technical and social aspects for the

70

research community, in practice, HR allocation techniques in software industry usually only respond to who does what. Assignments are typically made to tasks or roles individually and issues related to how individuals are assembled to form

75

M

the team are often disregarded. We propose here a general socio-technical framework to form teams globally. The motivation behind this framework is finding

ED

an optimal trade-off between technical and social aspects in order to have a successful project. The goal is to form a team of experts with the required technical knowledge, but also collaborating as a team in an effective manner. Our

80

PT

proposal benefits of some of the socio-technical aspects previously discussed in the research literature. During the evaluation of the technical dimension, we

CE

look for an optimal fitting among the technical needs and the candidates’ skills. During the evaluation of the social dimension, the aim is to build and analyze networks of software professionals characterizing and measuring their interper-

AC

sonal relationships. These social networks allow us to study the compatibility

85

among people who are involved in the project and identify those individuals that, working together, may potentially have an optimal joint-performance to complete the project. A relevant novelty of the proposed framework is that it does not only measure networks connecting software professionals, but also connecting the social skills they use to communicate, persuade or interact with other 5

ACCEPTED MANUSCRIPT

90

people (e.g., abilities such as verbal and non-verbal communication, empathy or leadership). Results obtained in multiple in-company experiments suggest that characterizing and measuring the interplay between coworkers when forming

CR IP T

software project teams can significantly improve the expected outcomes. We would like to emphasize that the proposed framework is general in the 95

sense that it can be apply to different kind of companies. The framework defines rules to evaluate and align relevant technical and social aspects affecting perfor-

mance of a software development team, but the information required to build

and measure the proposed social networks (e.g., project performance estimators

100

AN US

or level of knowledge/expertise in a given skill) can be obtained from different sources and/or be estimated using different metrics. This allows adapting the

framework to companies with different organizational structures and processes: from small companies to multinational corporations.

M

2. Related work

2.1. Socio-technical approaches to software development Interpersonal relationships are associations among two or more people formed

ED

105

in the context of a social interaction, e.g., friendship, affinity or communications. In this work, we are interested in collaborations between software professionals

PT

within a development team. Multiple types of personal and professional interpersonal relationships arise between coworkers in an IS project. Over the last 110

years, there is an increasing interest of these relationships on development tasks

CE

and literature on software engineering recognizes the importance of social aspects in the software development process (e.g., see Tamburri et al. (2013)). A

AC

non-trivial IS project can be considered a socio-technical endeavor where both technical and social factors influence the development process (Sawyer, 2004;

115

Luna-Reyes et al., 2005; Herbsleb, 2007). The technical component consists of the technical properties of the software to be developed, the development techniques or the technology employed. The social component includes the organization and the individuals involved in the development process, their attitudes

6

ACCEPTED MANUSCRIPT

and behaviors. Some authors claim that socio-technical congruence is a good in120

dicator of organizations performance in carrying out IS projects (Cataldo et al., 2006, 2008). Other investigations suggest that interpersonal relationships be-

CR IP T

tween coworkers determine the likelihood of development teams to be cohesive and productive (Yang and Tang, 2004; Adams and Anantatmula, 2010). 2.2. Social network analysis and the staffing problem

Socio-technical approaches to software development demand to assess the

125

relationships among coworkers in order to find compatible teammates during

AN US

the personnel assignment process. However, when putting together a team, it is difficult to determine who can efficiently work with whom. Compatibility can be given by technical (e.g., compatible programmers use a similar programming 130

style and naming strategy) and/or social aspects (e.g., fluidly communications and knowledge sharing between teammates).

Interpersonal relationships can be represented and analyzed with a social

M

network (Moreno, 1934), i.e., a graph where each node corresponds to an individual and links represent relationships between them. Social Network Analysis (SNA) has been applied in multiple scenarios from different scientific areas (Al-

ED

135

bert and Barab´ asi, 2002). In the context of software engineering, it emerges as a tool for capturing the underlying interaction processes within develop-

PT

ment teams and studying different aspects of the collaboration and organization among software professionals. Thus, SNA has been used, for instance, to iden140

tify top developers and their attitudes and knowledge sharing behaviors (Sowe

CE

et al., 2008; Licorish and MacDonell, 2014, 2015) or to predict failures and vulnerabilities in software products (Meneely et al., 2008; Wolf et al., 2009; Shin

AC

et al., 2011). SNA is also extensively used to analyze the Open Source Software development process (Crowston and Scozzi, 2002; Madey et al., 2002; Duche-

145

neaut, 2005; Bagozzi and Dholakia, 2006). More related to our work is the investigation by Lappas et al. Lappas et al.

(2009), who apply SNA to the problem of assigning people to a working task. Given a pool of candidates with different skills, a social network that captures 7

ACCEPTED MANUSCRIPT

the compatibility among these individuals and a particular task that requires 150

a set of skills, Lappas et al. (2009) addresses the problem of identifying the individuals best suited to perform the given task. Extensions of this algorithm

CR IP T

have been proposed later, for example, to be applied when a skill requires more than one expert (Li and Shan, 2010).

All these works provide a useful theoretical framework successfully tested 155

with scientific-collaboration graphs (Newman, 2001). Nevertheless, when you try to apply this framework to real IS projects in industry, you find some lim-

itations. The main issue is that Lappas et al. look for groups of individuals

AN US

that “together” have the required skills regardless of who displays each skill.

This is not a real scenario in industry, where projects demand individuals with 160

a specific set of skills. Furthermore, Lappas et al. focus on the network algorithms and do not address the construction of the social network capturing the compatibility among coworkers. In their tests, they build scientific-collaboration networks from the DBLP bibliography database where two scientists have a pre-

165

M

vious successful collaboration when they appear in the database as coauthors of at least one manuscript. In the following section, we review different common

ED

approaches to build social networks of software professionals. 2.3. Strategies to build social networks in software engineering

PT

Literature describes three main strategies to build networks representing social relationships among software professionals. The most common strategy is 170

based on the correlation observed between work performance and effective com-

CE

munications between coworkers (e.g., see Yang and Tang (2004)). Coworkers in an IS project can interact verbally and/or textually through different chan-

AC

nels. Investigations that use communications to identify relationships among coworkers found several problems in tracing and characterizing verbal commu-

175

nications, even when they were computer-mediated. Then, most of these works build collaboration networks using different heuristics to mine knowledge from mailing lists (Ducheneaut, 2005; Bird et al., 2006; Abreu and Premraj, 2009), bug tracking tools (Bhattacharya and Neamtiu, 2010; Bhattacharya et al., 2012) 8

ACCEPTED MANUSCRIPT

or version control systems (Bird et al., 2006; Schuler and Zimmermann, 2008; 180

Meneely and Williams, 2011; Bettenburg and Hassan, 2012), between others. Thus, they establish a connection between individuals that interchange a given

CR IP T

number of emails, commit the same file in a software repository or comment the same bug in a bug tracking system. Some studies show that relevant links

can be missing from the data in the social networks constructed with these 185

strategies (Bachmann et al., 2010).

A second strategy consists of using the hierarchical structure of the company

to identify the edges of the social network. This strategy has been used to

AN US

study how organizational structures interfere with task assignments and their influence on communications (Damian et al., 2013) and the resulting software 190

quality (Nagappan et al., 2008).

A last strategy consists of building social networks connecting participants in the same project (Madey et al., 2002; Xu et al., 2006). Note that these networks are equivalent to the scientific-collaboration networks mentioned above.

195

M

An interesting work where this strategy is used to find compatible programmers is Surian et al. (2011). In this work, they build collaboration networks con-

ED

taining three types of nodes – developers, projects and project properties – and define a metric to quantify the compatibility among developers as a function of the projects they have worked together before and their properties. The idea

200

PT

inspiring this metric is that two teammates working in the same project have common interests and skills. Surian et al. report “reasonable” results based on

CE

an input developer and snapshots from Sourceforge.Net. However, they do not address the construction of teams composed of people with different skills and

AC

roles.

205

These three strategies have been successfully applied to different relevant

problems in software engineering. Nevertheless, as we mention above, from our perspective, they have some limitations to be applied to the team allocation problem in industry. In addition to these limitations, we consider crucial to establish how positive or negative the interactions among coworkers are, and none of them does it. For instance, non-productive teammates can interchange 9

ACCEPTED MANUSCRIPT

210

a great amount of emails or systematically participate in failing projects. Social networks built in the context of the team formation problem must capture this information to avoid including individuals with negative relationships in future

CR IP T

working teams. To solve the drawbacks raised, we use a different approach. As Surian et al. (2011), we consider that the most reliable source for informa215

tion about compatibility between coworkers is their participation in previous projects, but we introduce a joint-performance estimator weighting each con-

nection in the social network (see Section 3.2 for details). Barreto et al. (2008) defines a similar “productivity modifier” characterizing the productivity of an

220

AN US

individual. The difference between this productivity score and ours is that Bar-

reto et al. assign the scores to individuals – nodes in terms of the social network – and we assign them to interpersonal relationships – i.e., connections. Moreover, we introduce an additional social score to analyze the collective compatibility of a group of software professionals, not only the compatibility between two

225

M

particular individuals.

3. Framework definition

ED

Our work departs from the hypothesis that the expected outcomes of new IS projects can be significantly improved if the project team is not only selected as a

PT

function of technical capabilities, but also considering interpersonal relationships within the team. Under this view, the goal of the proposed socio-technical 230

framework is maximizing the success probability of new teams finding an optimal

CE

trade-off between technical and social aspects. To achieve this goal, we build different compatibility social networks for employees and new candidates. The proposed framework is generic in the sense that it can be adapted to the needs

AC

of different type of companies. In particular, different strategies and metrics

235

can be used to construct and measure the different social networks proposed. Let assume a forming-storming-norming-performing team life cycle (Tuck-

man and Jensen, 1977), where a new team advances from the forming to the performing stage to become an effective team. The simple and intuitive idea behind

10

ED

M

AN US

CR IP T

ACCEPTED MANUSCRIPT

PT

Figure 1: Schematic representation of the proposed socio-technical framework.

our proposal is putting together compatible teammates with the required technical skill in order to facilitate this transition and minimize the storming stage

CE

240

impact. By compatible, here we mean people that working together reach a high joint-performance. To predict the most adequate project team composition, the

AC

framework analyzes multiple variables. Depending on the company, this includes different project metrics, the employees’ technical and non-technical capabili-

245

ties, and/or the social interactions within the team, between others. To briefly provide an intuitive view of this analysis, Fig. 1 illustrates schematically the proposed framework:

11

ACCEPTED MANUSCRIPT

1. First, we identify the candidates best suited to participate in the project (Section 3.1.1). When possible, these will be the members of the different teams to evaluate and the nodes of our social networks. Suitability is

250

CR IP T

determined through the matching between the attitudes required for each task or role (Pi , identified with a different color in Fig. 1) and the available human resources’ attitudes.

2. Connections in the social networks are established as a function of previous collaborations, i.e., projects where the candidates worked together before.

255

Then, we extract the information of similar projects from a database of

AN US

completed projects (Section 3.1.2).

3. Once potential candidates and projects to rate their previous collaborations have been found, we build two kind of compatibility networks. On one hand, social networks characterizing the interplay between team-

260

mates in terms of previous productive and non-productive relationships (Section 3.2.1). On the other hand, social networks connecting candidates

M

with compatible and incompatible social skills (Section 3.2.2). Individuals in the social-skill networks can be connected even when they have not worked together before.

ED

265

4. Finally, the potential candidates are combined using different strategies (Section 3.3), and the resulting teams are analyzed with SNA to estimate

PT

the most effective assignments for new projects or to include new members in existing teams (Section 3.4).

As we mention above, this general schema can be adapted to the specific

CE

270

context of the company where the framework is applied. In the following sections, we describe possible strategies to construct and measure our teammate

AC

and social-skill networks. 3.1. Identification of nodes and links

275

3.1.1. Find potential candidates Nodes of our social networks are the potential candidates to participate in the project. Like in any other strategy for assigning people to a working task, 12

ACCEPTED MANUSCRIPT

potential candidates are identified as a function of the fitting among the candidate’s abilities and the set of required abilities. To group these abilities, we use 280

the concept of profile – defined as the set of specific attitudes for each task or

CR IP T

role in the project. A profile can include different type of attributes: technical and non-technical skills, personality traits, limiting individual factors (e.g., trav-

eling availability) and personal preferences. Note that this requires identifying

and grading every candidate’s technical and non-technical skills and personality 285

traits (see threats to validity in Section 6). While some existing decision support

systems for assigning people to IS project teams already incorporate the candi-

AN US

dates’ technical and non-technical skills (Andr´e et al., 2011; e Silva and Costa,

2013), considering limiting individual factors and personal preferences is a novelty of our proposal. The need of taking into account limiting individual fac290

tors is obvious for real-life industrial projects. Regarding personal preferences, experience indicates that when an employee is continuously assigned to tasks he/she does not like to perform and is qualified to perform a higher-level job,

M

at the end, a decreasing performance likely takes place. Thus, a possible sample of web developer’s profile is Pdeveloper = {java, TDD, german competencies, stress tolerance, not traveling availability, distributed project}.

ED

295

In the simpler case, profiles only include technical skills, since this is the minimal information to identify those candidates with the required technical

PT

knowledge. The inclusion of other elements is not mandatory and depends on the company and the information it manages. To match the required profiles and the available human resources’ profiles,

CE

300

these are considered fuzzy sets in order to manage uncompleted associations between the required and the actual level of knowledge/expertise in a given skill

AC

or capability (Drigas et al., 2004). When there is not a perfect match between profiles and candidates, we use the best-fitted resource (BFR) methodology

305

proposed in Otero et al. (2009). BFR analysis quantifies how prior knowledge in other skills contributes to the learning of a given skill. This analysis allows us to build effectively teams even when the optimal profiles are not completely available. 13

ACCEPTED MANUSCRIPT

3.1.2. Find similar projects We measure networks built considering the projects where individuals have

310

worked before. Multiple variables make an IS project to be unique. Therefore,

CR IP T

the generation of our social networks is driven only by data of similar projects to the one for which the new team is being formed. Similarity between projects is

estimated using a clustering algorithm. These algorithms allow us to compute 315

and compare similarity vectors between projects. In our implementations we perform a k-means clustering analysis (MacQueen, 1967; Xu and Wunsch, 2005). The identification of similar projects extensively depends on the data stored

AN US

in the corresponding database of projects (see threats to validity in Section 6).

Results reported in this paper consider vectors with the following features to 320

determine similarity among projects: type of project, area of knowledge, working methodology, estimated workload, complexity, risk level, team size, budget and geographical distribution. We would like to highlight the latter, since geographical distribution usually is disregarded during the project team compo-

involves teams split over locations and countries.

ED

325

M

sition (beyond language competencies). However, current software production

3.2. Building compatibility networks 3.2.1. Teammate networks

PT

A teammate network – e.g., see Fig. 2 – is social network that consists of nodes representing individuals (IT M = {I1 , ..., In }) and links connecting them and quantifying their expected compatibility with a compatibility score (RT M =

CE

330

{CT MIi Ij }). The more compatible a teammate-teammate relationship in the network, the higher the estimated joint-performance of the corresponding indi-

AC

viduals. Links between teammates are bidirectional and CT MIi Ij = CT MIj Ii . If two individuals are not connected in the network, we have no evidence re-

335

garding their compatibility. This is different from being incompatible. Although the previous definition is conceptually simple, it is not easy to

determine a priori the compatibility among coworkers. Similar approaches to ours described in the related work assume that two individuals are compatible 14

AN US

CR IP T

ACCEPTED MANUSCRIPT

Figure 2: Example of teammate network. Scores in the links indicate how compatible the corresponding interpersonal relationship is. For simplicity in the graphical representation,

M

positive relationships are plotted in green color; and negative in red.

and have a successful collaboration simply if they previously worked in the same project. From our perspective, this and other similar heuristics do not capture

ED

340

the compatibility among coworkers in a real-life scenario, because they do not measure whether interpersonal relationship were actually positive. To solve

PT

these drawbacks, we rate connections in our teammate networks using two types of metrics: subjective metrics based on the teammates’ personal opinion about their interactions, and objective metrics based on the collective performance

CE

345

of previous teams. Thus, subjective teammate networks are constructed as a function of a subjective metric; and objective teammate networks as a function

AC

of an objective metric.

350

We would like to emphasize again that the proposed framework is defined

to be applicable in different contexts. In this regard, the compatibility score of a teammate-teammate relationship can be given by different subjective or objective metrics depending on the particular company. It can be expressed

15

ACCEPTED MANUSCRIPT

qualitatively or quantitatively, and obviously, the possible values and the interpretation of these values vary depending on the strategy and metrics used 355

to build the corresponding teammate network. The only relevant point is the

CR IP T

order relation between the possible values. In the case of quantitative data, we assume that the higher value corresponds to the more compatible relationships.

If this is not the case, it can be easily changed. To simplify the analysis, qualitative data are converted into numerical values maintaining the order relation. 360

For example, in a “positive-negative-neutral” Likert scale, we can assign 1 to “negative”, 2 to “neutral” and 3 to “positive”.

AN US

Subjective teammate networks are based on the idea that only teammates

know how they interact during daily work. Then, connections in these networks are rated according to the assessment the teammates themselves supply. 365

Depending on the context of the company, different strategies can be used to achieve this. The HR management domain provides us with different tools to evaluate interactions among coworkers. In particular, we highlight here the so-

M

called 360-degree feedback (London and Beatty, 1993), since this is one of the mechanisms most frequently used for development and assessment purposes in multiple organizations. Briefly, 360-degree surveys gather feedback on an indi-

ED

370

vidual from different sources, typically including self-evaluation, subordinates, peers and supervisors. Subjects have to answer questions rating a partner in

PT

different topics related to team climate and team dynamics (e.g., leadership, teamwork, decision-making, job knowledge or productivity, between others). Then, statistical analysis can be used to obtain comprehensive performance and

CE

375

behavioral evaluations of each individual. When some of these appraisal systems is implemented in the organization, they usually provide us with information

AC

we can use for building subjective teammate networks. Another mechanism to collect the information required to build subjective teammate networks is using

380

one of the questionnaires described in Appendix B. This a simple manner to obtain these data when they are not available in the company, as well as when the existing assessment process requires too much effort from the employees (see threats to validity). 16

M

AN US

CR IP T

ACCEPTED MANUSCRIPT

Figure 3: Simple illustrative example of teammate network (left panel) and social-skill network (right panel) built based on the two projects on top. In this example, the networks are gener-

ED

ated considering the normalized effort deviation as team performance estimator. As in Fig. 1, green and red connections correspond to positive and negative relationships respectively.

385

PT

Unlike subjective networks, objective ones are not built on the basis of the employees’ opinion, but on an objective evaluation of their previous collaborations. Thus, objective metrics are joint-performance estimators that assume

CE

that members of performing teams are compatible and have positive relationships among them; while failing projects have non-compatible teammates with

AC

negative relationships. This requires a global project performance/success es-

390

timator, being the score of a particular link the sum of this estimator in all the projects where the corresponding teammates worked together (Fig. 3, left panel). The advantage of this strategy is that this kind of metrics is quite common in IS companies. Multiple metrics have been discussed in software engineering literature for project success, e.g., on-time/on-budget completion, 17

ACCEPTED MANUSCRIPT

395

system quality or effort/budget deviation (DeLone and McLean, 2003; Espinosa et al., 2006). The proposed framework adapts to the information managed by the company, and teammate networks can be based on any of these metrics (see

CR IP T

threats to validity in Section 6). Finally, we also test teammate networks that combine subjective and ob400

jective metrics. The motivation behind these networks is to identify negative

personal relationships among teammates even when they reach a high level of accuracy when working together and vice versa. To calculate the score of the

we use the following equation:

AN US

connection between two individuals (Ii and Ij ) in these collaboration networks,

CT MIi Ij = sig0.25 (CT MIsi Ij ) · sig3.0 (CT MIoi Ij ) · CT MIoi Ij 405

(1)

where sigβ (x) = 1/(1 + eβx ) is the sigmoid function, and being CT MIsi Ij and CT MIoi Ij the normalized subjective and objective score of the corresponding

M

link, respectively.

In the results section, we compare outcomes obtained with teammate net-

410

ED

works built with different strategies and metrics. 3.2.2. Social-skill networks

The motivation behind the social-skill network is the existence of particular

PT

combinations of social skills – i.e., personality traits and abilities a person uses to interact with other people – that make two people “connect” during team-

CE

work and have a productive professional relationship (Blaskovich, 2008; Adams 415

and Anantatmula, 2010; Feldt et al., 2010). In this situation, the assessment of relationships among coworkers as a function of their social skills may help to

AC

determine the optimal project team composition, in particular when they have not previously worked together (e.g., consider companies with a high turnover rate). Note that in this situation a teammate networks do not provide informa-

420

tion regarding compatibility. Different works in literature investigate the personality traits present in exceptional software engineers (Turley and Bieman, 1995), in the individuals that 18

ACCEPTED MANUSCRIPT

dominate team communication (Licorish and MacDonell, 2015) or the ones required to perform a specific development activity (Acu˜ na and Juristo, 2004; 425

Acu˜ na et al., 2006). However, we do not found definitive results pointing out

CR IP T

which are the relevant skills to be included in our social-skill network. Therefore, along with a group of occupational psychologists belonging to the HR

department of one of our industrial partners, we defined a list of 11 social skills

to be measured in the social profile of an individual: written communication, 430

verbal and non-verbal communication, active listening, teamwork, negotiation,

sociability, empathy, agreeableness, leadership. Appendix A contains a brief

AN US

description of what we mean by each of these social skills. The intention of

this 11-skill social profile is to determine and predict each candidate’s ability to communicate and interact with coworkers avoiding or minimizing conflicts 435

within the working team (see threats to validity).

The question now is how to determine an individual’s social profile. Again, the HR management domain provides us with tools to determine the level of

M

development of an individual in different social skills, including the ones of our social profile. For example, the Myers-Briggs Test Indicator (MBTI), the Big Five (BF) and the Sixteen Personality Factor Questionnaire (16PF) are com-

ED

440

monly used in software engineering literature (e.g., see Hardiman (1997); Acu˜ na and Juristo (2004); Feldt et al. (2010); Andr´e et al. (2011); Licorish and Mac-

PT

Donell (2015)). These are simple personality tests nowadays readily available online which we use in different in-company experiments. In addition to these tests, multiple organizations implement their own competencies assessment sys-

CE

445

tem to identify their employee’s personality traits. Depending on the company,

AC

the framework can receive the social profiles from these or other sources. A social-skill network is equivalent to a teammate network, but reflect-

ing the correlation in terms of productivity between different skills (Fig. 2,

450

right panel). Nodes in these networks represent skill-level of development pairs (SSK = {S1 , ..., Sn }). We consider as possible levels of development for the different social skills, as possible values in the corresponding metric used to assess candidates in this skill. Once identified the nodes in the network, to 19

ACCEPTED MANUSCRIPT

establish connections between them, we depart from an objective teammate 455

network. Two nodes (Si and Sj ) are connected when this network contains a link between individuals with the corresponding skill-level of development pairs.

CR IP T

The compatibility score of the link (CSKSi Sj ) is then calculated as the sum of the scores of all the connections between individuals with these skills in the teammate network. Social-skill networks do not directly measure compatibil460

ity among candidates. They serve to compute an objective metric (C(I1 , I2 )),

which can be used to build a new teammate network (IT M 0 = {I1 , ..., In }; RT M 0 = {CT MI0i Ij = C(I1 , I2 )}) reflecting the compatibility among poten-

AN US

tial candidates from the perspective of their social profiles. Teammate networks built in this way can be measured as any other objective or subjective team465

mate network (Section 3.4). Formally, given the social profiles of two individuals (SPI1 = {s1I1 , ..., snI1 } and SPI2 = {s1I2 , ..., snI2 }) and a social-skill network (SSK = {S1 , ..., Sn }; RSK = {CSKSi Sj }), C(I1 , I2 ) is estimated as follows: XX

M

C(I1 , I2 ) =

i

j

CSKsi

sj I1 I2

(2)

As an example, let consider the following social profiles {negotiation: high; 470

ED

leadership: high; listening: medium} and {negotiation: low; leadership: low; listening: low}. Given the social-skill network depicted in the right panel of

0.2.

PT

Fig. 3, the compatibility between two individuals with these social profiles is

CE

3.3. Strategies for defining potential teams for the project In our socio-technical approach, the technical dimension is evaluated in a

475

first stage to identify potential teams for the project. Note that we use the

AC

term technical in a broad sense, since our profiles can include technical and non-technical skills, personality traits, limiting individual factors and personal preferences (Section 3.1.1). To do this initial identification, we use a BFR methodology to match the required profiles and the available resources’ profiles

480

(see Section 3.1.1). Multiple combinations of candidates can satisfy the project technical requirements producing a great amount of potential teams, especially 20

ACCEPTED MANUSCRIPT

when the number of available resources is high. Then, SNA can be applied to every potential team to assess the social dimension (Section 3.4). However, we can use algorithms based on our social networks to limit the number of assessment and drive the direction of our predictions. 3.3.1. Individuals compatibility strategy

CR IP T

485

The individuals compatibility strategy departs from the premise that people

who have successfully worked together before will work successfully again. Note

that this does not imply maintaining permanent teams composed by the same people. In the context of the proposed socio-technical framework, working suc-

AN US

490

cessfully means having a high compatible score in the teammate network. Then, we only generate teams composed of individuals included in the project’s teammate network. If a compatible teammate is not found for any of the required profiles, additional team members are assigned considering the project’s social495

skill network (see Section 3.3.2). These second assignments are made taking

included in the same team.

M

into account that incompatible teammates in the teammate network are never

ED

Our results indicate that the individuals compatibility strategy is the best strategy to substitute a key member (e.g., the team leader) in already existing 500

teams once a project has begun.

PT

3.3.2. Social-skills compatibility strategy The social-skills compatibility strategy consists of creating the team finding

CE

compatible individuals as a function of their social skills, i.e., considering the information provided by the project’s social-skill network. In production, this

505

strategy is mainly used to create teams including people who have never worked

AC

together, e.g., employees recently hired. 3.3.3. Star-based strategy Some works in literature suggest that people with highly developed social

skills can facilitate project success (Licorish and MacDonell, 2014, 2015). The 510

proposed teammate networks allow building teams based on this idea. If we 21

ACCEPTED MANUSCRIPT

compute the Katz centrality (Katz, 1953; Grindrod et al., 2011; Sun and Tang, 2011) of every node in the teammate network of a project, we can identify individuals that normally interplay effectively with their coworkers. We call

515

CR IP T

these individuals social-stars. The social-star-based strategy consists of building the team around some of these individuals. Once the social-stars are selected, the team is completed using the social-skill network.

Similarly to the social-star-based strategy, the team can be formed around

individuals highly skilled to perform a given role or task. In this case, the technical-stars are selected as a function of the project technical needs and the team is also completed using the social-skill network.

AN US

520

3.3.4. Split strategy

The split strategy consists of splitting a performing team in two new teams. New teams are completed with individuals with the same or similar social skills to the ones observed in the members of the original team. This strategy is mainly used with groups of programmers or testers with the idea of increasing

M

525

the pool of effective relationships.

ED

3.4. Analyzing compatibility networks To mine knowledge from the proposed social networks, we compute two measures for each potential team (T ). On one hand, the number of positive links within the team (P RT ). Depending on the strategy and the metric used to con-

PT

530

struct the network, positive has a different interpretation. For example, in the

CE

case of a subjective teammate network based on 5-point qualitative metric that rates relationships between teammates as “not effective”, “somewhat effective”, “effective”, “very effective” and “extremely effective”, we will consider positive those links labeled as “effective”, “very effective” or “extremely effective”. On

AC

535

the other hand, we compute the team social score (T SST ), defined as the cu-

mulative compatibility score of the relationships between every team member within the team. Formally, if T is composed by the teammates {tm1 , ..., tmm }: XX T SST = CT Mtmi tmj (3) i

j

22

ACCEPTED MANUSCRIPT

As the compatibility scores computed for links belonging to two distinct 540

teammate networks, team scores derived from different networks are not comparable. In this sense, it allows comparing teams in the context of a given social

CR IP T

network, i.e., teams assessed under the same conditions. The greater the social score of a team, the higher its estimated success probability.

Once computed these two values, the goal is to find an optimal balance be545

tween maximizing T SS and P RT . Let consider a simple illustrative example

departing from the teammate network depicted in Fig. 2. Our goal is to create a new 4-member team with a unique common required profile, which, for simplic-

AN US

ity in the example, all subjects in the network have. SNA allows us to compute easily the T SS value for all the possible teams: 26 out of 70 possible 4-member 550

teams do not have negative relationships, being T1 = {I4 , I6 , I7 , I8 } the team with the greatest T SS (T SST1 = 17). Then, only considering T SS, the network indicates that T1 is the best team to be assigned to the project. Nevertheless, P RT1 = 2, i.e., we only have evidence of 2 out 6 positive relationships among

555

M

the members of T1 . If we now consider T2 = {I1 , I3 , I4 , I6 }, T SST2 = 16 and P RT2 = 5. This is also a good recommendation from the team social score

ED

viewpoint (cf. T SST1 and T SST2 ), but the teammate network provides additional evidence of a good joint-performance between all members of T2 except I3 and I6 . Furthermore, assuming transitive relationships between software

560

PT

professionals, i.e., relationships between collaborators of collaborators (Surian et al., 2011), we could predict that I3 and I6 will be compatible too, because

CE

both are compatible with I1 and I4 . Transitive relationships of a given individual are computed by the inspection with depth 2 of the nodes accessible from the corresponding node in the network. Then, considering the information pro-

AC

vided by the network of Fig. 2, T2 is the 4-member team with a greatest success

565

probability.

23

ACCEPTED MANUSCRIPT

4. Proposal validation To validate our proposal, we carried out different kind of experiments in our industrial partners. In these experiments, we addressed two main goals. First,

570

CR IP T

to verify our proposal applicability and the suggested teams coherence. Second,

to validate whether the proposed teams were actually effective and productive. For statistical analysis, we group our industrial partners into four categories

depending on the number of employees. According to the definitions currently in

used in the European Union1 , companies with less than 50 and 250 employees are considered small- and medium-sized companies, respectively; while large-sized

enterprises have more than 250 employees. Although some of the medium- and

AN US

575

large-sized companies have employees located in offices from different countries, we consider an additional category (multinational organization) for more-than1000-employee enterprises with teams split over multiple countries.

M

4.1. Simulation of team assignments

To look into the applicability and coherence of our proposal in a given com-

580

pany, we first performed simulations of team assignments. These are proof-of-

ED

concepts in which we define a series of required profiles for fake projects and use the framework to form the corresponding team with the company’s current employees. Depending on the company, required profiles are obtained from the planning of previously completed projects and/or they are randomly chosen from

PT

585

a pool of predefined profiles. Then, a group of project managers, team leaders

CE

and/or HR management professionals designated by the company validates the assignments. When they reject a team, they must indicate the rejection causes. Before running the simulations of team assignments, the framework must

be adapted to the company. The aims of this initial phase are to identify the

AC

590

information already available in the company that can be used to build the social 1 EUR-Lex:

96/280/EC: Commission Recommendation of 3 April 1996 concerning the

definition of small and medium-sized enterprises, Official Journal L 107, 30/04/1996 P. 0004 - 0009

24

ACCEPTED MANUSCRIPT

networks, and define the protocols to capture required data. The company’s HR department usually participated in the framework adaptation. In many cases, they proposed elements to incorporate into the social networks – e.g., additional competencies and/or social skills they include in the employee’s profile or specific

CR IP T

595

metrics provided by the tools they use to assess team climate. To quantify the

adaptability and applicability of the framework, we calculate the percentage of companies where simulations could be run after the initial adaptation phase.

Once the framework is adapted to the company, simulations of team assign600

ments allow us to compare coherence of the teams proposed by social networks

AN US

generated with different metrics and strategies. Depending on the data available in the company, we can construct, measure and compare:

• Subjective teammate networks generated as a function of the employees’ responses to questionnaires Q1 and Q2 of Appendix B. 605

• Subjective teammate networks built with the information provided by any

M

existing tool to evaluate coworkers’ collaborations.

• Objective teammate networks generated using different project metrics.

ED

For statistical analysis, we distinguish between four success/performance indicators. Effort and budget deviation – i.e., the difference among es610

timated and actual effort/budget – since these are the most commonly

PT

available indicators in our industrial partners. Another project metrics, such as customer satisfaction or different quality measures, are only avail-

CE

able in a few companies. We group all these metrics into a unique category in order to get enough statistical evidence. Finally, a last simple

AC

615

metric, which can be even used when the company does not keep track of any project metric, consists of rating the projects as “successful” or “unsuccessful” depending on whether they are completed on schedule and meeting budget and requirements.

• Teammate networks combining different objective and subjective metrics 620

(Eq. 1). 25

ACCEPTED MANUSCRIPT

• Social-skill networks built with the subjects’ social profiles. Coherence of the teams obtained from networks constructed with different strategies is evaluated in terms of the percentage of accepted teams. These

625

CR IP T

values are compared with results obtained in control experiments where team

members are selected randomly using two possible strategies. With the first strategy, all the team members are selected randomly from the pool of available

resources independently of their skills. With the second one, each member is selected randomly too, but now from the set of candidates best suited for each

630

AN US

task or role. 4.2. Pilot experiments

Simulations of team assignments allow us to verify the applicability of our proposal and the coherence of the suggested teams, but they do not validate whether these teams are actually effective. To do this, we carried out pilot

635

M

experiments where the proposed framework served as a tool to help decisionmakers allocate the team of real IS projects. In the lead-up to these experiments, we always performed simulations of team assignments to adapt the framework

scenario.

ED

to the company and identify the social networks providing better results in each

In the pilot experiments, the final decision about the team composition is on the managers’ hands. Then, as in the simulations, we study the percentage

PT

640

of teams suggested by our social networks and assigned to real projects. We

CE

use this value to quantify the managers’ confidence in the framework. Note the difference between accepting a team in a simulation and assigning this team to a production project in the pilot experiments. In addition to this implicit measure of the managers’ perception, we also analyze what they think the framework

AC

645

brings to the team allocation process. For that, they have to fill in with answers a questionnaire with the following questions: • QPE1. How helpful do you think the tool for selecting the project team is? (1-5) 26

ACCEPTED MANUSCRIPT

• QPE2. I identified suited people for the projects who I had not taken into

650

consideration (1-5)

CR IP T

• QPE3. The tool allows forming more efficient project teams (1-5) • QPE4. The tool could be used to form project teams automatically (1-5) The questions are expressed on a Likert scale, where 1-5 mean “Not at 655

all helpful”, “Slightly helpful”, “Moderately helpful”, “Very helpful” and “Ex-

tremely helpful” for QPE1; and “Strongly disagree”, “Disagree”, “Neither agree nor disagree”, “Agree” and “Strongly agree” for QPE2-4.

AN US

Once a team is assigned to a production project, we verify whether this assignment was effective by evaluating four common project productivity metrics: 660

a binary success indicator, effort deviation, budget deviation and customer satisfaction. In a study like ours, it is nearly impossible to quantify to what extent the project benefited of a specific team composition as compared to other as-

in the same real project.

4.3. Large-scale production studies

ED

665

M

signments, since we cannot compare results obtained by different teams working

After the pilot experiments, some companies adopted the proposed framework as a supporting tool helping decision-makers during the creation of teams

PT

for IS projects. This demonstrates their confidence in our socio-technical approach, but moreover the large-scale production studies allow us to verify improvements in the team allocation process (see threats to validity).

CE

670

Although as in the pilot experiments, we cannot compare results of different

teams undertaken the same project, in the large-scale studies we do have enough

AC

statistical sample to compare performance of teams formed using the proposed framework and teams formed using other strategies (the specific strategy de-

675

pends on the company and the project). Comparisons are made in terms of different measures, mainly productivity and quality metrics (e.g., effort/budget deviation or customer satisfaction), depending on the company and the expected benefits. To minimize possible bias, results of the teams suggested by 27

ACCEPTED MANUSCRIPT

Personality traits (Yes-No)

Collaborations (Yes-No)

Applicability

Small-sized

9-6

2-13

0-15

100%

Medium-sized

16-3

10-9

8-11

Large-sized

5-0

4-1

2-3

Multinational

2-0

2-0

2-0

TOTAL

41

41

41

CR IP T

Project success (Yes-No)

Type of company

89.5% 100%

100%

95.1%

Table 1: Summary of the companies where the framework was tested. Middle columns indicate

AN US

the number of companies with the corresponding measures already available before performing

the simulations of assignments. The last column shows the percentage of companies where the framework was successfully applied.

the framework are compared with baselines established averaging the results of 680

projects completed by teams formed with the allocation process that we want to

M

improve. An improvement in the metrics analyzed in the large-scale production studies does not guarantee the assignment of the most effective team, but it gave us enough statistical evidence to verify that the framework has a positive

5. Results

PT

685

ED

impact on the company.

5.1. Applicability of the proposed framework

CE

To date, we have simulated team assignments in 41 organizations of different size and with different internal processes and hierarchical structure (Table 1). This has allowed us to validate the proposed framework adaptability and applicability to multiple scenarios. The framework was successfully tested in 39

AC

690

out 41 (95%) of our industrial partners. In these 39 companies, we were able to capture the required data and/or adapt the framework to the data previously managed by the company and run the simulations. The two companies where the simulations failed had quite similar proper-

695

ties: around 100 software professionals involved in the project, the only informa28

ACCEPTED MANUSCRIPT

tion available of these subjects was their curricular data, and historical data of projects were not available. In both companies, we found two problems during the framework adaptation. First, it was nearly impossible to build a reliable

700

CR IP T

base of knowledge of successful and unsuccessful previous teams. Second, the data acquisition process required some effort on the company: employees must

perform different tests to identify their personality, behavioral and social traits;

and they must complete the questionnaires of Appendix B to assess their interpersonal relationships with coworkers. Due to these issues, the project was

canceled before running the simulations. In this regard, it must be said that data capture was a bottleneck in some medium- and large-sized companies be-

AN US

705

cause of the number of questionnaires/evaluations each employee had to perform before running the simulations (see threats to validity). In spite of this, simulations were successfully run in six medium and one large enterprise with similar characteristics to the companies where they were canceled (i.e., without pre710

vious data regarding personality traits and collaborations). Four of these six

M

medium-sized companies are currently using the framework as a supporting tool

ED

to form the production teams of their IS projects. 5.2. Teams coherence

The simulations of team assignments indicate that the proposed social networks in general suggest coherent project teams (Fig. 4). No matter the type

PT

715

of company or the strategy used to construct the network, the experts designated by the company consider around 90% of the suggested teams coherent

CE

and valid for the project (cf. control experiments). The greater acceptance percentage occurs for small-sized companies due to the small number of potential candidates. For instance, acceptance is 100% in the three companies with less

AC

720

than 12 employees. However, this is not a significant result. Projects in these companies often required very homogeneous development profiles (e.g., android developer with knowledge in JSON, XML, Java and SQL), and most of the employees had these skills. This makes nearly all the possible candidates to be

725

considered adequate for the project. This observation is corroborated by the 29

CR IP T

ACCEPTED MANUSCRIPT

Figure 4: Acceptance percentage of the teams suggested by our teammates and social-skill networks in the simulations of team assignments. Results of the control experiments are

grouped into teams randomly selected from the whole set of available resources (“Random”)

AN US

and the ones selected from the set of candidates best suited for each task or role (“Random best”). Objective and subjective teammate networks are grouped into different categories depending on the metric used to build the network. “HR tool” corresponds to networks based on data extracted from an existing tool (note that none of the small-sized companies has this kind of tool). “Success indicator” corresponds to a binary successful/unsuccessful metric. “Other” includes project metrics but effort and budget deviation. See Section 4.1 for details.

M

control experiments, where teams selected randomly in small companies have a high acceptance percentage (88% or 96% depending on the strategy used to

ED

select the team members). Therefore, we cannot draw any conclusion about coherence of the teams suggested by the framework in small-sized companies. 730

We discuss below the helpfulness in these organizations.

PT

Comparing the different strategies to generate teammate networks, teams proposed by objective networks have in general a better acceptance than the

CE

proposed by subjective networks. This occurs regardless of the objective team performance estimator or the mechanisms used to assess subjectively interper-

735

sonal relationships (Fig. 4). Although it seems obvious that only the team

AC

members themselves have a detailed knowledge of the social interactions in the project, the simulations suggest that objective joint-performance estimators have a greater accuracy to capture efficiency of these interactions, and, therefore, objective networks contain more relevant information for us. A more

740

detailed analysis corroborates this observation. For example, teams formed with

30

AN US

CR IP T

ACCEPTED MANUSCRIPT

Figure 5: Objective and a subjective teammate network generated for a medium-sized company with projects completed within the last five years. For simplicity in the visualization, we only show connection for one employee (represented in blue in both networks). The jointperformance indicator used to build the objective network is the binary success indicator, i.e.,

M

successful projects compute +1 and failing projects −1. The subjective network is based on the last-five-year data extracted from an assessment tool already existing in the company. It scores the relationships between employees with a value between 0 and 100. The greater

ED

the score, the more positive the interpersonal relationship (ranges 0-39, 40-60 and 61-100 correspond to negative, neutral and positive relationships, respectively). Colors in both cases are used to identify positive, negative and neutral relationships according to the information

PT

captured in the corresponding network.

the social-stars strategy have a greater acceptance when social-stars are iden-

CE

tified in objective teammate networks. These networks also allow us to detect positive relationships between individuals with limited social skills that do not appear in networks generated with subjective metrics. For example, the blue individual of Fig. 5 represents a senior software developer. In the last five years,

AC

745

he participated in 11 project, six completed successfully and five unsuccessfully. His social profile indicates that he has very limited social abilities: excellent and acceptable competence in written communication and leadership, respectively; and a poor or very poor level of development in the rest of social skills. This

750

translates into a difficulty interacting with teammates, which is reflected in the 31

ACCEPTED MANUSCRIPT

subjective teammate network (Fig. 5, right). Taking into account this network, this employee was never proposed in the simulations as member of a new team. However, while the network built with subjective metrics does not show any

755

CR IP T

positive connection, six productive interactions can be observed in the equivalent objective network (Fig. 5, left). In view of this network, for example, a

team containing the blue individual, I3, I6 and I7 will be likely productive (cf. connections in the right panel). If the large-scale production experiments con-

firm that teams established in this way are actually productive, this is a highly relevant property for the team assignment problem in the context of a complex

collaborative process like software development. The objective teammate net-

AN US

760

works would allow identifying teammates able to adapt their working styles to special circumstances as the difficulty in social interactions.

A possible explanation for the problems found in the assignments derived from subjective evaluation of interpersonal relationships is that these metrics 765

are biased by other positive or negative relationships, e.g., affinity or friendship,

M

which do not always translate into productive or non-productive professional relationships. It is important to highlight that this bias appears in the simple

ED

questionnaires proposed in Appendix B, but also in a well-accepted HR assessment of team climate like 360-degree feedback. In fact, our results indicate that 770

the teams suggested by networks generated using questionnaires Q1 and Q2, and

PT

the teams suggested by a HR-tool-based network have nearly the same components and a very similar acceptability (Fig. 4). This result points out that, for

CE

our needs, the simple questionnaires of Appendix B collect equivalent information to the feedback gathered by more complex assessment processes. In the

775

context of a social-technical approach like ours, this is a highly relevant result

AC

for companies without corporate mechanisms to assess collaborations and team dynamics. Note that none small companies, and only 8 out of 19 medium-sized implement this kind of tool.

780

Concerning results of the different objective teammate networks, one could

expect that the higher the accuracy of the success/performance estimator, the more reliable the corresponding network. Nevertheless, the stakeholders’ general 32

ACCEPTED MANUSCRIPT

perception is that the optimal results are obtained when the networks are built as a function of the effort deviation. Note that this metric largely depend on the estimation quality – e.g., if a project fails due to over-optimistic scheduling or under-staffing, the corresponding team is considered non-productive.

CR IP T

785

Despite the possible bias discussed above in the subjective teammate net-

works, our analysis points out that subjective evaluation of collaborations can add some value when it is incorporated as correction factor to objective networks. Networks built in this way slightly improve the percentage of acceptance 790

of objective teammates networks (Fig. 4). Studying the teams proposed by these

AN US

networks, we observe that the information provided by teammates about their

interpersonal relationships allows measuring personal preferences together with the productive dimension of professional collaborations. Given how the compatibility score is computed in these networks (Eq. 1), productive relationships 795

where both teammates think that the collaboration is positive (although biased by friendship) are considered “more productive”. In the same way, subjective

M

evaluation makes “more negative” those non-productive relationships between teammates with a non-good personal relationship.

800

ED

Finally, results of the social-skill networks are nearly the same to the obtained with the teammate networks (in all cases they have a greater than 90% acceptance). It is important to remind that the teams proposed by these net-

PT

works are not based on previous collaborations, but on the estimated compatibility between their members’ social skills. The larger the company size, the

CE

more coherent the teams proposed. This result corroborates our departing hy805

pothesis that putting together coworkers as a function of their social skills is a good strategy to build IS project teams (at least from the coherence point of

AC

view).

5.3. Managers’ perception

810

We performed pilot experiments in 28 industrial IS enterprises. Table 2

summarizes the properties and measures used to build the proposed social networks in each company. It is important to remind that these are fixed after a 33

ACCEPTED MANUSCRIPT

simulation phase (see Section 4.2).

2

small (19)

3*

small (44)

4

small (24)

5*

small (41)

6*

small (45)

7

small (25)

8

small (30)

9

small (32)

10*

medium (87)

11*

medium (213)

12*

medium (232)

14*

ED

15

medium (105)

16*

medium (91)

17*

medium (203)

18

medium (64)

19

medium (133)

20*

medium (215)

21*

medium (152)

medium (123) medium (99)

AC

CE

PT

13*

effort deviation effort deviation defect density success indicator success indicator success indicator success indicator effort deviation defect density effort deviation defect density effort deviation defect density success indicator success indicator budget deviation budget deviation success indicator effort deviation effort deviation effort deviation

Q2 Q2

Personality traits

34

#teams

CR IP T

small (48)

Subjective measure

existing HR tool

#accepted

27

26

MBTI-BF-16PF existing HR tool

10

10

21

20

MBTI-BF-16PF

17

15

MBTI-BF-16PF

36

32



MBTI-BF-16PF

34

32

Q1

MBTI-BF-16PF

14

11

Q1

MBTI-BF-16PF

12

11

Q1

MBTI-BF-16PF

6

6



MBTI-BF-16PF

40

35



MBTI-BF-16PF existing HR tool

52

40

82

74

MBTI-BF-16PF existing HR tool existing HR tool

25

22

35

26

35

28

MBTI-BF-16PF existing HR tool

41

29

67

58

MBTI-BF-16PF existing HR tool existing HR tool existing HR tool

28

27

24

21

79

71

56

47

Q2 Q2 Q1

AN US

1*

Objective measure

M

Company

Type of company (participants)

Q1 Q2 existing HR tool Q2 existing HR tool existing HR tool Q2 — existing HR tool existing HR tool

ACCEPTED MANUSCRIPT

medium (111)

23

medium (55)

24*

large (289)

25

large (325)

26*

large (312)

27

large (450)

28*

multinational (1819)

TOTAL

effort deviation effort deviation effort deviation effort deviation success indicator budget deviation effort deviation

Subjective measure

Personality traits

#teams

#accepted



MBTI-BF-16PF

21

17

CR IP T

22*

Objective measure

Q1 existing HR tool Q2 Q2

Q1 existing HR tool

MBTI-BF-16PF existing HR tool existing HR tool

MBTI-BF-16PF existing HR tool existing HR tool

AN US

Company

Type of company (participants)

6

6

88

71

56

43

109

94

71

63

380

340

1472

1275

Table 2: Features of the pilot experiments. Note that, in line with the results discussed in Section 5.2, we always measured team-

M

mates networks built based on objective metrics, or on the combination of objective and subjective metrics. Companies identified

ED

with (*) incorporated the proposed framework to their staffing pro-

PT

tocol and/or are currently performing large-scale studies.

We first analyze the project managers’ opinion about the framework. Table 3 shows the distribution of their answers to the questionnaire regarding the helpfulness of the framework during team assignments (Section 4.2). In general,

CE 815

they think that the framework helps decision-making, allowing the identifica-

AC

tion of employees suited for the project that they did not know or, at least, they did not initially consider for the project. In spite of this, they believe that the team allocation must be supervised (cf. response to QPE4). Beyond the

820

general result, there are some significant differences in the managers’ perception of the framework depending on the company size. A significant number of managers belonging to small companies do not consider the framework useful

35

2

3

4

Small-sized

40.1%

23.2%

2.8%

20.3%

Medium-sized

1.4%

5.7%

9.5%

60.4%

Large-sized

2.2%

11.1%

6.5%

40.4%

Multinational

0.8%

1.6%

8.4%

38.1%

TOTAL

6.1%

8.0%

7.7%

45.5%

Small-sized

51.4%

10.2%

2.3%

27.1%

Medium-sized

1.2%

3.9%

1.9%

46.7%

46.3%

Large-sized

0.3%

4.6%

7.1%

42.9%

45.1%

Multinational

1.3%

2.4%

3.2%

52.4%

40.7%

TOTAL

7.1%

4.4%

3.4%

45.0%

40.1%

Small-sized

50.3%

7.3%

24.9%

16.9%

0.6%

Medium-sized

1.9%

2.9%

10.8%

69.7%

14.7%

Large-sized

6.8%

1.2%

9.9%

34.0%

48.1%

Multinational

0.0%

1.3%

9.5%

39.5%

49.7%

TOTAL

8.3%

2.6%

12.0%

47.7%

29.4%

Small-sized

68.9%

24.3%

6.8%

0.0%

0.0%

PT

5

Medium-sized

53.1%

34.0%

8.5%

4.4%

0.0%

Large-sized

92.3%

7.7%

0.0%

0.0%

0.0%

Multinational

97.9%

0.5%

1.6%

0.0%

0.0%

TOTAL

75.2%

18.4%

4.6%

1.8%

0.0%

CR IP T

1

CE

QPE4

13.6%

23.0%

39.8%

51.1%

32.8% 9.0%

AN US

M

ED

QPE3

QPE2

QPE1

ACCEPTED MANUSCRIPT

AC

Table 3: Distribution of the managers’ responses to the questions about the framework helpfulness. They are expressed on a Likert-type scale as follows. QPE1: 1 = Not at all helpful; 2 = Slightly helpful; 3 = Moderately helpful; 4 = Very helpful; and 5 = Extremely helpful. QPE2-4: 1 = Strongly disagree; 2 = Disagree; 3 = Neither agree nor disagree; 4 = Agree; and 5 = Strongly agree. See Section 4.2 for details.

36

ACCEPTED MANUSCRIPT

(40.1% not at all helpful; and 23.2% slightly helpful). They believe that they can make team assignments as a function of their experience and knowledge 825

about the potential team members, and the framework does not provide them

CR IP T

with any additional information. This is in line with the results obtained in the simulations of team assignments, and can be something expected in companies where everyone knows each other (cf. response to QPE2). Interestingly, all the managers of small companies who consider the framework very helpful or 830

extremely helpful belong to more-than-40-employee companies. As the number of potential candidates grows, this perception disappears and managers tend to

AN US

think that the proposed social networks help them understand the contribution of the people they manage.

In addition to the managers’ opinion, we also analyze their confidence in the 835

framework by means of the percentage of teams accepted and assigned to real projects. Figure 6 shows that this value is high (around 85%, see also column “#accepted” in Table 3), but always lower than the percentage of accepted

M

teams in the simulations. The analysis of the rejection causes indicates that this is because, in production, managers tend to reject the teams composed of people that do not have worked together yet. Although in the simulations they

ED

840

consider these teams coherent and valid for the project, in production they prefer not to do these assignments when a team composed of previous teammates can

PT

be formed. As some of the suggested teams successfully advances, confidence in the framework assignments grows, and a greater number of teams based on the social-skill networks is accepted and assigned to new projects. Note that,

CE

845

in spite of the initial doubts, the mean acceptance is around 87% (86% if we do

AC

not take into account the small-sized companies). 5.4. Team effectiveness validation

850

Figure 7 shows the outcomes of the teams assigned to real IS projects from

different perspectives. We first consider that a project is successfully completed when the system is delivered on-time, within-budget and meets the required features and functions (“success indicator”). Likely, this is the indicator that best 37

CR IP T

ACCEPTED MANUSCRIPT

Figure 6: Percentage of acceptance of the teams suggested by our teammates and social-skill

CE

PT

ED

M

AN US

networks in the pilot experiments.

AC

Figure 7: Outcomes of the project teams suggested by the socio-technical framework.

summarizes the team success. From this perspective, 87.9% of the teams can

be considered successful teams. Fails from this perspective are mainly a conse-

855

quence of scheduling deviations (∼ 83% of the failing projects). Interestingly, only 6% of the failing teams do it because of the system quality. Results from the effort, budget and customer satisfaction viewpoint are nearly equivalent, al-

38

ACCEPTED MANUSCRIPT

though we would like to highlight results from the effort deviation perspective. In some projects, the team needs more time than the estimated to complete 860

some tasks. This does not always translate into a budget deviation or a quality

the effort deviation are the more reliable.

CR IP T

decrease, but it can explain why the teammate networks built as a function of

In view of the results shown in Fig. 7, we can conclude that the teams suggested by the framework are in general effective and efficient. 865

5.5. An empirical study

AN US

To date, 17 out of 39 companies (Table 2) where the framework was successfully applied are currently running large-scale production studies, i.e., 44% of the companies (41% considering all the companies where we tried to apply the

framework; 54% if we do not include companies with less than 20 employees; and 870

61% if we only consider those companies that performed pilot experiments). The most relevant empirical study is ongoing at a multinational corporation based

M

in Europe, Asia and South America (company 28 of Table 2). After performing different simulations and pilot experiments with employees and projects belong-

875

ED

ing to different countries, the company decided to address the improvement of the people assignment process in nine software factories that showed a general deviation in scheduling and budget. They were located in Spain (2), France (3),

PT

Germany (1), Mexico (1) and Argentina (2). In this section, we focus on the results of measuring the proposed social networks when assigning people to the projects faced by these nine software factories during more than three years. Appendix C describes the steps leading up to the framework application.

CE 880

AC

5.5.1. Effectiveness and efficiency improvements The expected benefits of applying the framework in the nine software fac-

tories were mainly related to effectiveness and efficiency enhancements. Then, we compare the productivity and quality results of teams formed before and

885

after the framework start-up. In particular, the analyses are made in terms of effort deviation (productivity) and defects per KLOC (quality). Baselines for

39

Individuals

Social-skills

Social-stars

Tech-stars

Split

Effort deviation

0.84 beall

0.81 beall

0.85 beall

0.85 beall

0.84 beall

0.83 beall

Defect density

0.77 bdall

0.75 bdall

0.79 bdall

0.79 bdall

0.76 bdall

0.77 bdall

Effort deviation

0.85 bees

0.82 bees

0.85 bees

0.85 bees

0.87 bees

0.86 bees

Defect density

0.79 bdes

0.77 bdes

0.83 bdes

0.81 bdes

0.78 bdes

0.79 bdes

Effort deviation

0.81 bef r

0.77 bef r

0.83 bef r

0.85 bef r

0.80 bef r

0.81 bef r

Defect density

0.76 bdf r

0.73 bdf r

0.79 bdf r

0.76 bdf r

0.78 bdf r

0.75 bdf r

Effort deviation

0.89 bede

0.90 bede

0.89 bede

0.87 bede

0.91 bede

0.88 bede

Defect density

0.79 bdde

0.75 bdde

0.81 bdde

0.84 bdde

0.74 bdde

0.79 bdde

Effort deviation

0.82 bemx

Defect density

0.73 bdmx

Effort deviation

0.81 bear

Defect density

0.79 bdar

CR IP T

Overall

AN US

ar

mx

de

fr

es

all

ACCEPTED MANUSCRIPT

0.77 bemx

0.84 bemx

0.86 bemx

0.83 bemx

0.81 bemx

0.71 bdmx

0.70 bdmx

0.71 bdmx

0.77 bdmx

0.75 bdmx

0.78 bear

0.84 bear

0.83 bear

0.80 bear

0.82 bear

0.79 bdar

0.84 bdar

0.84 bdar

0.72 bdar

0.77 bdar

Table 4: Productivity and quality measures of the on-site projects undertaken by teams

M

formed with the support of the proposed framework. Table 5 displays equivalent measures for distributed projects. Results are shown as a function of the three-year baselines for effort deviation (bex ) and defect density (bdx ); globally (“all”) and grouped by country (“es”,

ED

“fr”, “de”, “mx” and “ar” correspond to Spain, France, Germany, Mexico and Argentine, respectively). Columns display results grouped by the strategy followed to form the team.

PT

comparison are computed with the projects completed within the three years prior to start applying the framework (see Appendix C). Table 4 shows the

CE

outcomes obtained in the on-site projects – i.e., projects where all the team 890

members belong to the same software factory – completed by teams formed with the framework support during a more-than-three-year period. Analyzing

AC

the overall results, we observe a general enhancement, both in productivity and quality. The mean effort deviation is 16% fewer; while the mean defect density has decreased 23%. Improvements by country are coherent with these values.

895

Teams composed of teammates who have successfully worked together before

generally are more effective (cf. the individuals compatibility strategy and the rest of strategies in Table 4). Most members of these teams know each other and 40

ACCEPTED MANUSCRIPT

have close professional relationships. The existence of prior positive relationships within the team favors collaboration and coordination and facilitates be900

coming a performing team. Conversely, the teams formed using other strategies

CR IP T

need time to grow and consolidate the teammate-teammate relationships. This implies that these teams follow a more pronounced forming-storming-norming-

performing pattern (cf. red and blue traces in Fig. 8). However, in spite of the

more pronounced forming, storming and norming phases, they usually become 905

performing teams after a relative short formative period. Note that the mean

effort deviation in Table 4 is similar independently of the strategy followed to

AN US

select the project team. This result is observed even in teams created using

the social-skills compatibility strategy, which does not take into account previous collaborations to evaluate relationships among coworkers. This observation 910

served to decide using this strategy to assign people to small non-critical projects in order to create new teammate-teammate relationships. These new relationships usually become effective during the project, and then the new compatible

M

teammates were assigned to projects that were more critical. In this regard, we would like to emphasize the high turnover rate: 26%, 24.5%, 19.5%, 28.5% and 17% in the last year in the Spanish, French, German, Mexican and Argentinean

ED

915

software factories. In this situation, it seems critical the formation of effective relationships among the employees recently hired.

PT

Although the statistical sample is enough to minimize bias, the improvements could be consequence of other factors, e.g., a more skilled employee recruiting. Furthermore, effort deviation and defect density are simple and in some

CE

920

sense unreliable measures of productivity and quality. Therefore, we would like to emphasize the stakeholders’ general perception. They believe that the use

AC

of the proposed social networks led to a noteworthy improvement in the new teams’ efficiency, which minimized the scheduling and budget problems in the

925

nine software factories. In fact, they have started three new large-scale production studies: one involving three software factories in India; another with five software factories in Mexico and five in Argentina; and the last one with 15 software factories located in Europe (3 in Spain, 8 in France and 4 in Germany). 41

CR IP T

ACCEPTED MANUSCRIPT

AN US

Figure 8: Evolution of the mean normalized performance of 20 different on-site French teams assigned to a more-than-one-year project. Red trace corresponds to 20 teams created using the individuals compatibility strategy. Blue trace corresponds to teams created using the social-skills compatibility strategy (7), the social-stars strategy (7) and the split strategy (6). The analysis covers the 12 months after the team allocation and its assignment to a new project. Dashed line denotes the French three-year performance baseline.

M

5.5.2. Emphasizing the relevance of social interactions

Beyond the general results shown in Table 4, additional observations after

930

ED

measuring the proposed social networks during more than 36 months can be summarized in the following points: • Results of teams created with the social-stars and the technical-stars strat935

PT

egy are nearly the same. This emphasizes the relevance of interpersonal relationships for project success and suggests that social skills can be as

CE

important as technical skills. In some cases, the social-stars strategy can produce teams with a better mean performance (see Spanish and Ger-

AC

man teams in Table 4). In step with these results, multiple projects were

940

successfully completed by development teams not having all the required technical skills, but having a high social score and a large amount of positive interpersonal relationships between their members. A detailed analysis indicates that, although from a technical viewpoint these were not good team assignments, the projects met their budget and quality ob-

42

ACCEPTED MANUSCRIPT

jectives due to an effective knowledge sharing. As expected, there are fluid 945

collaborations within the working team, but also with employees outside the team. The proposed social score is not defined to predict this kind of

CR IP T

collaborative behavior. However, our analysis points out that the greater the number of social-stars in a project team, the higher the knowledge sharing outside the team. 950

• Taking into account the results of Licorish and MacDonell (2014, 2015), the social-starts strategy was initially proposed thinking in people in man-

agement, architectural and development roles. Nevertheless, depending

AN US

on the country, around 27-39% of the individuals identified as social-stars perform an ancillary role. These results point out again the relevance 955

of the social aspects. Despite that, we would like to highlight the IS project leaders’ profile. It is well-known that project leaders play a crucial role and can be a major source of errors that lead to failure due to

M

bad decision-making (Charette, 2005). However, we have found that relationships between project leaders and the rest of team members are also 960

critical for project success. Multiple non-effective teams become effective

ED

under the command of project leaders with social skills compatible with the ones of the team members. Note that we do not only refer to specific

PT

project leaders, but to different leaders with specific social-skills. • Teammates with limited social skills (e.g., people showing a low team965

work capacity) can also constitute high-performing teams. These effective

CE

relationships usually appear between highly skilled technicians that, in general, only have productive relationships with similar peers. This kind

AC

of teams successfully performs small projects or specific – and usually

970

complex – tasks. If this is not the case, the team needs complementary members with additional social skills. • When a performing team has a high maturity level, replace some of its members does not significantly affect the team’s effective productivity. In this regard, we would like to stress results obtained with the split strategy 43

ACCEPTED MANUSCRIPT

since, departing from a performing team; it permits obtaining two nearly 975

equivalent teams to the original one. This strategy increases the number of productive teammate-teammate relationships.

CR IP T

• Finally, teammates and social-skill networks have helped to create project

teams distributed across different software factories. This was not a common practice in the company. Then, initially, distributed teams were 980

created using the social-skills compatibility strategy. A great percentage of these initial teams failed, even when teammates belonged to the same

country. This suggests the need of different social skills for interacting

AN US

with distributed coworkers. Once the number of productive distributed teammate-teammate relationships grew, performing teams were formed 985

using the individuals compatibility strategy. Afterwards, the increased number of productive relationships allows successfully using other assignment strategies. For instance, high-performing teams have been created

M

around distributed social-stars identified in the teammate networks. Currently, distributed teams’ mean performance and quality are nearly the 990

same than the on-site teams (cf. Tables 4 and 5). Although it is difficult

ED

to draw general conclusions due to the individual-dependent nature of social interaction, distributed teammate networks allow us to analyze how

PT

compatible social profiles change depending on the teammate’s location. For example, around 87% of people that successfully interact with cowork-

995

ers of other countries show a high empathy level. Similarly, more than

CE

75% of these people also have a medium-high adaptability level. In other cases, a dependence on the language used to communicate is observed.

AC

For example, around 72% of the effective teammates who communicate in

1000

English show a high active listening level. There even exist skills needed to interact successfully with coworkers of specific countries.

44

ACCEPTED MANUSCRIPT

Spanish

English

Effort deviation

0.84 beall

0.83 besp

0.87 been

Defect density

0.84 bdall

0.85 bdsp

0.84 bden

CR IP T

Overall

Table 5: Outcomes of the distributed teams composed of people belonging to different software

factories. Columns labeled as “Spanish” and “English” indicate the language used to communicate. Communication is in Spanish language if the project involves people from Spain,

Mexico and/or Argentina; and it is in English when teammates are from Spain, France and/or Germany.

AN US

6. Threats to Validity

To limit the scope of our claims as well as to explain their utility, the first important fact to take into account is that this is an empirical study. Like most empirical studies, the validity of our results is subject to several threats (Basili 1005

et al., 1999).

M

Concerning the generalization of the findings, we obtained successful outcomes in companies of different size, with different organizational structures and different internal processes. Nevertheless, results presented in this paper

1010

ED

should not be taken as valid in the general sense. Although a Hawthorne effect is not possible, the environment in real industrial projects cannot be controlled to the same extent as in isolated research experiments. In this regard, it is impor-

PT

tant to note that each project is different and that team dynamics are complex. These factors make nearly impossible to control all the variables affecting suc-

CE

cess or failure of an IS project. Thus, there exist several threats concerning

1015

external factors that may affect our results. In particular, we do not exactly know whether the performance and quality improvements observed in the pi-

AC

lot experiments and the large-scale production studies can be attributed to the selection of more adequate coworkers or to many other possible factors. Furthermore, obtaining optimal results in the metrics analyzed in the pilot experiments

1020

and the large-scale production studies does not guarantee the assignment of the most effective team in a general sense. In the first case, our results must be in-

45

ACCEPTED MANUSCRIPT

terpreted as the formation of effective teams from parameters established by the company in each particular case. In the second case, they must be interpreted as an improvement in team assignment process of the company. Regarding the statistical conclusion validity, we analyze results derived from

CR IP T

1025

tests and pilot experiments performed in 39 different companies. This includes

thousands of simulations and real project assignments. Furthermore, we report detailed results of using the proposed framework during more than three years

in nine software factories belonging to a multinational corporation. This gives 1030

us enough statistics to consider that our results are statistically valid.

AN US

We assume that the most reliable source for information about compati-

bility between coworkers is their common participation in previous projects. In particular, we define compatibility as a function of objective global metrics related to success/performance of completed projects. This definition is also 1035

subject to some threats to validity. With this approach, we cannot guarantee that some relevant nodes and/or links from the network can be missing from the

M

source data, as it happens with other strategies for constructing compatibility networks for software professionals (Bachmann et al., 2010). If the database

1040

ED

of previous projects is not complete, multiple nodes or links can be missing. Similarly, predictions of the social networks also depend on the reliability of the metrics managed by the company. Depending on the particular data managed

PT

by the company, in our implementations we mainly used the following variables to measure project success/performance: on-time completion, within-budget

CE

completion, budget/effort deviation, system quality and customer satisfaction. 1045

These are the variables identified in literature as the main metrics for global IS projects success (DeLone and McLean, 2003; Espinosa et al., 2006). However,

AC

we also obtained successful outcomes with other metrics such lines of code per unit of effort or time taken to complete a task. Finally, with our definition, we only analyze intra-team relationships, when relationships among team members

1050

and the rest of the organization can also be relevant. In spite of these threats, our empirical results demonstrate that the abstraction that we propose is a good approach for the team allocation problem. 46

ACCEPTED MANUSCRIPT

Regarding the adaptability to different organizations, the proposed social networks can be built using data and metrics commonly managed by indus1055

trial companies. The employees’ technical and non-technical capabilities can be

CR IP T

captured from curricular data, which are commonly available in industrial companies. Personality traits and interpersonal relationships can be assessed with

different metrics. These metrics include estimators from the HR management domain that can be obtained from tools commonly available in medium/large1060

sized enterprises – e.g., 360-degree feedback surveys (London and Beatty, 1993) or competencies assessment (Robertson and Smith, 2001). If the company does

AN US

not have any of these tools, MBTI, BF and 16PF are simple personality tests

nowadays readily available online. These three tests are commonly used for personnel assessment in software engineering literature (e.g., see Hardiman (1997); 1065

Acu˜ na and Juristo (2004); Feldt et al. (2010); Andr´e et al. (2011); Licorish and MacDonell (2015)) and we successfully used them in some of our experiments. In this regard, we strongly recommend dealing with the help of professionals of

M

the organization’s HR department, since they know the mechanisms existing in the company to assess employees and their collaborations. This aims to adapt the framework to the specific characteristic of the company. For instance, in

ED

1070

some of our in-company experiments, the HR department proposed additional competencies measured by their own personality tests to be included in the

PT

social networks.

Concerning the applicability of the proposed framework, one of the main drawbacks is related to the data required to build and measure the social net-

CE

1075

works. It needs specific data related to projects (e.g., features, participants and success/performance metrics) and employees (e.g., detailed profiles or estima-

AC

tors of their social abilities). This need can make the framework not useful in some organizations. On one hand, although as we indicate above it can be

1080

easily adapted to the data managed by each company, not all the companies manage and keep track of the information required. In these cases, not only the team selection process must be implemented, but also additional processes to capture and keep historical data of projects and employees. On the other 47

ACCEPTED MANUSCRIPT

hand, the larger the company, the greater the volume of data to be captured 1085

and analyzed. In this regard, it is important to note that, although subjective data about working interpersonal relationships among coworkers can contribute

CR IP T

to improve teamwork, this information is not mandatory for the framework. In fact, objective teammate networks generated with a, in some sense, unreliable productivity measure as the effort deviation, provide more relevant information 1090

than those built with data subjective data.

7. Discussion

AN US

Many researchers and practitioners in software engineering hold to a socio-

technical view of IS projects (Sawyer, 2004; Luna-Reyes et al., 2005; Herbsleb, 2007; Tamburri et al., 2013). From this view, both the technical and social 1095

aspects of a project are crucial to performance and, therefore, should be taken into account to succeed. While multiple mechanisms to evaluate and improve

M

the technical dimension of a project are extensively applied in software industry, the mechanisms to control and manage the social dimension are still unclear. The framework presented in this paper takes advantage of some previously discussed elements and incorporates novel principles for globally assigning people

ED

1100

to IS projects from a socio-technical perspective. We propose the assessment of

PT

the candidates’ technical skills, experience and personality traits, but also their interpersonal relationships. For that, the framework leans on the information managed by the company to create personal profiles of each employee and to construct and measure different social networks which objectively and/or sub-

CE

1105

jectively characterize the collaborations among coworkers. This characterization can be made according to previous (successful or unsuccessful) collaborations

AC

as well as to a measure of the compatibility between teammates estimated as a function of their social skills. For companies where the required information

1110

is not available, we define possible mechanisms to collect it in a simple manner. We provide industrial insights into the framework applicability in different contexts, obtaining successful outcomes according to the parameters established

48

ACCEPTED MANUSCRIPT

by the corresponding company and with social networks constructed and measured with distinct strategies and metrics, which points out the generality of 1115

our proposal.

CR IP T

How to build efficient and effective teams is one of the most relevant questions in software engineering management. In practice, the decision of who participates in a project normally is on the managers’ hands. Although we also

found companies where the assignments were made by the HR department as 1120

a function of curricular data. Managers usually make decisions based on their

experience and subjective knowledge about the potential team members. In

AN US

our in-company experiments, we observed that in this scenario decision makers

tended to count on the same people and many new projects directly inherited teams of previously completed projects. This practice, that might be evident in 1125

small companies with a limited staff, also happened in medium- and large-sized companies and even in cases of recurrent failing teams. In many occasions, this was due, on one hand, to the lack of knowledge sharing among managers and, on

M

the other hand, to the lack of proper methods to evaluate personnel capabilities. Experience showed that maintaining teams composed by the same people was not always an effective practice, as successful teams for a given project might

ED

1130

not be appropriate for another one. Moreover, in the organizations with more complex structures, managers could be part of many different teams, which

PT

could be distributed across different locations or even countries. In these cases, they did not know every coworker daily work and did not always fully understand the contribution of a particular member to the team and his/her impact

CE

1135

on team dynamics. In these cases, the same as when they were made by the HR department, assignments were often made as a function of the candidates’ cur-

AC

ricular data. Therefore, we found a strong argument for a tool like the proposed framework, defining mechanisms to obtain wide-ranging information to form an

1140

accurate picture of coworkers’ performance and their interactions and, leaning on this information, infer adequate configurations for a specific project. In this regard, experiments in real-life industrial projects pointed out that managers, in general, appreciated the existence of such mechanisms, but they preferred to re49

ACCEPTED MANUSCRIPT

ceive the information and define their owns teams according to their experience 1145

(Table 3). The teammate networks provided managers with information that allowed

CR IP T

them to identify potential effective relationships among employees they did not know. They were also suitable to point to groups of teammates that usually work together in failing projects. In these networks, the team social score max1150

imizes the former relationships and minimizes the latter within new working teams. The pilot experiments and the long-term production studies showed

that forming teams based on this value is an effective strategy to create per-

AN US

forming teams. Since social-skill networks are not based on previous collaborations, the compatibility information they provided for decision makers was 1155

more difficult to understand than the one provided by the teammate networks. At a first glance, the managers were reluctant to accept these relationships and form teams based on them. However, the experiments also demonstrated the general good performance of these teams (cf. results of the individuals and the

1160

M

social-skills compatibility strategy in Table 4). From a more theoretical perspective, results obtained from the social-skill networks support the thesis that

ED

specific combinations of social skills allow two teammates to collaborate more effectively in an IS project. In this regard, we would like to highlight the particular case of some employees with limited social capabilities having difficulty

1165

PT

interacting with others (e.g., see Fig. 5). The analysis of our social networks permitted to identify coworkers with compatible social skills and the ability to

CE

adapt their working styles to these individuals. Beyond the benefits of forming more performing teams, these team assignments also increased the general personnel satisfaction. In addition to the direct benefits of creating more effective

AC

working teams, some of the staffing strategies implemented using the proposed

1170

social networks, e.g., the split strategy, had relevant positive medium/long-term benefits in exchange for a little short-term performance penalty (Fig. 8). This strategy is conceived to generate two performing teams from one existing performing team. In our experiments this was usually achieved (Table 4). Then, new productive teammate-teammate relationships provided an increased capac50

ACCEPTED MANUSCRIPT

1175

ity to face successfully new projects for the company. This is a highly valuable result for organizations with a high turnover rate. On the down side, the information required to build the proposed social

CR IP T

networks. To implement the framework in an industrial enterprise, we need: • A detailed knowledge of each employee’s technical skills and capabilities is mandatory to find potential candidates to participate in the project.

1180

• Access to historical productivity/quality metrics of each project is mandatory to build objective teammate networks.

AN US

• To take advantage of the social-skill networks, it is needed to know each employee’s social profile. This is likely the information more difficult to collect.

1185

Beyond the use of the proposed framework, our results emphasize the importance of how participants in an IS project collaborate, communicate and, in

M

general, work together as a team. In line with previous findings (Fitzpatrick and Askin, 2005; Acu˜ na et al., 2008; Adams and Anantatmula, 2010; Feldt et al., 2010), this points out that the social aspects can be as relevant as the technical

ED

1190

aspects for successfully completing a project (cf. results of the social-start and the technical-start strategy in Table 4), and supports the hypothesis that we can

PT

significantly improve our outcomes paying attention to interpersonal relationships when putting together a team. Social factors may enhance (or diminish) 1195

the sum of individual team members’ attitudes. In this sense, for instance,

CE

we found individuals with a low productivity that significantly increased their performance when moved to a team with different teammates; or projects suc-

AC

cessfully completed by teams not covering all the required technical skills but composed of people with compatible social skills. Obviously, people are not the

1200

only factor to be considered, and we cannot disregard other critical aspects for the project success such as the planning, project tasks, workload or scheduling, between others. The benefits of considering software development in industry as a socio51

ACCEPTED MANUSCRIPT

technical process were observed independently of the company size. Never1205

theless, they were less apparent in small-sized enterprises where the reduced number of employees, and usually also of projects, limited the number of possi-

CR IP T

ble assignments. In this regard, it is important to highlight that implementing a framework like ours to find an optimal trade-off between technical and so-

cial aspects does not have sense in this kind of companies. In addition to the 1210

limitation in the number of possible teams, managers in these companies often

know the rest of employees during daily work (Table 3). Although we have not enough statistics to raise a general conclusion, throughout our experiments we

AN US

established a threshold of 20 software professional to consider the framework helpful.

We have developed social networks characterizing interpersonal relationships

1215

within the team. These relationships are strongly related to group factors such as cohesion, conflicts or team climate, between others. However, social factors in an IS project are not limited to intra-team aspects. Of course, interpersonal re-

1220

M

lationships outside the working team can significantly impact on performance. Our teammate networks could incorporate additional elements characterizing

ED

interpersonal relationships outside project teams. For instance, Oh et al. (2006) found a correlation between team effectiveness and what they called the group social capital – the set of resources available to the team through its members

1225

PT

interpersonal relationships within the team and with other members of the organization. In companies with the ability to compute the group social capital

CE

proposed by Oh et al., it should be possible to develop a strategy for team assignments trying to balance this value and our team social score. Given the current trend towards globalization in the software development

AC

industry, we encourage software organizations to systematically incorporate re-

1230

lationships between coworkers to their team staffing processes. In this scenario, software production involves teams split over locations, countries, and even companies. Despite the increasing attention to distributed software projects, there is a limited understanding of team dynamics that makes these projects succeed or fail. In a distributed development team, there are many additional factors 52

ACCEPTED MANUSCRIPT

1235

influencing social interactions (Holmstrom et al., 2006; Espinosa et al., 2006). Between others, these include temporal distance (i.e., people working at different time zones which implies few overlapping hours), socio-cultural differences

CR IP T

(e.g., native language and/or culture differences) and geographic distance. Due to these communication issues, in a distributed project there exist less commu1240

nications among coworkers: each individual communicate with fewer coworkers

and interactions are less frequent (Olson and Olson, 2000). Even some au-

thors claim that communications are less effective (Herbsleb and Mockus, 2003). Therefore, distributed software engineering poses a challenge to the development

1245

AN US

process, demanding additional behavioral traits and new social abilities in the

interaction among software professionals. Results reported in this paper in the context of distributed projects are a good departing point to incorporate SNA to the team formation problem in software industry.

file

M

Appendix A. Proposal of competencies to include in the social pro-

This Appendix briefly describes the competencies and social skills measured

1250

ED

in our social-skill networks. As we indicate in the main text, these data can be obtained from different tools depending on the company. This implies that in different social-skills network the same value has different meanings and inter-

1255

PT

pretations.

• Written communication: Ability to write clearly and effectively in order

CE

to express information via a written channel

AC

• Verbal and non-verbal communication: Ability to give information and

1260

actively manage the communication process through oral channels.

• Active listening: Ability to accurately receive and interpret messages in the verbal communication process. • Teamwork: Ability to contribute to teams and improve their effectiveness through personal commitment. 53

ACCEPTED MANUSCRIPT

• Negotiation: Ability to talk or consult together with someone for the purpose of mutual arrangement using a win-win philosophy. • Sociability: Ability to get on well with a wide range of people and build

1265

CR IP T

long term trusting relationships.

• Empathy: Ability to see the world as another person, to share and understand another person’s feelings, needs, concerns and/or emotional state.

• Agreeableness: Ability to be kind, sympathetic, cooperative, warm and

AN US

considerate with other.

1270

• Leadership: Ability to guide and inspire others towards achieving goals. • Motivation: Ability to support and encourage individuals, so that they give their best.

• Adaptability: Ability to respond and adapt to changing circumstances and to manage, solve problems and provide solutions.

M

1275

ED

Appendix B. Questionnaires to evaluate teammates collaborations Information required to built subjective teammate networks can be extracted from an appraisal system previously implemented in the company. However,

1280

PT

these tools require of consistent organizational processes and they are not available in many companies. Furthermore, they usually require too many effort

CE

from subjects (e.g., in some of the appraisals already existing in our industrial partners each individual has to complete questionnaires with more than 200 questions). Then, in order to simplify the process, a group of occupational psy-

AC

chologists of one of our partners helped us define two possible questionnaites to

1285

collect subjective information regarding the teammates’ collaborations in a simple manner (Table B.6). One of the goals of the simulations of team assignments (Section 4.1) is comparing the results obtained with social networks generated on the basis of information provided by some existing tool in the company, and the ones obtained with questionnaires Q1 and Q2. 54

ACCEPTED MANUSCRIPT

ID

Question

Q1

What is the impact of Ana on your work? (1-5)

Questionnaire Q2

CR IP T

Questionnaire Q1

Question

Q1

How productive is your work with Ana? (0-10)

Q2

How effective is your work with Ana? (0-10)

Q3

How well does Ana communicate with you? (0-10)

Q4

How well does Ana collaborate with you? (0-10)

Q5

How well does Ana help you with your work? (0-10)

Q6

How respectfully does Ana treat you? (0-10)

Q7

How quickly does Ana respond your requests? (0-10)

Q8

How well does Ana handle criticism? (0-10)

Q9

How professionally does Ana behave? (0-10)

M

AN US

ID

ED

Table B.6: Questionnaires to collect subjective evaluations of the teammates’ collaborations.

Questionnaire Q1 was defined to capture with a single question whether a

1290

PT

subject considers positive, negative or neutral his/her professional relationship with a particular teammate (Ana in the example). Using a common strategy in this kind of evaluations, the only question of Q1 is expressed on a five-point

CE

Likert-type scale (1 = Extremely negative; 3 = Neither positive nor negative;

1295

and 5 = Extremely positive). In questionnaire Q2, subjects have to rate nine

AC

questions with a numerical value between 0 and 10. These questions refer to well-known key aspects in social interactions within a work team – e.g., communication or knowledge sharing (Lievens et al., 2002; Chen et al., 2004).

1300

When the network is built leaning on one of these questionnaires, connections

among potential candidates are scored with the mean value of the responds provided by the two candidates about their collaboration.

55

ACCEPTED MANUSCRIPT

Appendix C. Background of the empirical study Appendix C.1. Framework adaptation

1305

CR IP T

We made the framework adaptation along with professionals of the Spanish HR department. The company kept track, between others, of the following rel-

evant data for all the projects undertaken by the software factories included in

the study: type of project, area of knowledge, working team, working methodology, estimated workload, complexity, risk level, budget, effort/budget deviation, customer satisfaction, defect density and software complexity. We used some of

these data (type of project, area of knowledge, working methodology, estimated

AN US

1310

workload, complexity, risk level, team size, budget and geographical distribution) to generate the similarity vectors employed to characterize and compare projects. While effort/budget deviation, customer satisfaction, defect density and software complexity were used to generate and test objective teammate net1315

works based on different metrics. For the simulations and the pilot experiments,

M

we generated a base of knowledge with all the projects completed within the last three years in the nine software factories. Once the large-scale production

ED

study started up, this was weekly updated. Baselines for comparison in these experiments were also computed with the initial last-three-year data. These 1320

baselines did not have changed along the study.

PT

The company also implemented at organizational level two tools whose data were suited to construct our social networks: a 360-degree feedback appraisal and a competencies assessment system. Every year, each employee has to pro-

CE

vide self-evaluation and assess a group of coworkers including his/her supervisor,

1325

subordinates and some colleagues. Feedback collected of an individual in the

AC

360-degree assessment is weighted to calculate overall individual and collective metrics measuring different aspect related to team work. During the framework adaptation phase, we decided to test the use of two of these metrics. On one hand, a 0-to-10 quantitative value rating teammate-teammate relationships: 0

1330

corresponds to relationships with negative impact on team dynamics and 10 to relationships between coworkers with an extremely positive interplay, respec-

56

ACCEPTED MANUSCRIPT

tively. This metric was used to test 360-feedback-based teammate networks. On the other hand, a qualitative score that ranges from 0 (not developed at all) and 10 (highly developed) the level of development of an individual in different competencies which could be linked with the skills of our social profile (Appendix

CR IP T

1335

A). Competencies assessment, for its part, supports new employees recruitment. Between others, it assesses with a 0-to-10 quantitative value 11 competencies

that also could be used to characterize an employee’s social profile. When this information was not available for an employee included in the study, he/she had to complete the competencies assessment process.

AN US

1340

Appendix C.2. Results of the simulations and the pilot experiments During the simulations of team assignments leading up to the framework application in the nine software factories selected by the company, we tested the following social networks:

• Subjective teammate networks based on Q1, Q2 and 360-feedback.

M

1345

• Objective teammate networks based on effort/budget deviation, customer

ED

satisfaction, defect density and software complexity. • Teammate networks based on all the possible combinations of subjective

PT

and objectives metrics.

• Social-skill networks based on the information extracted from the existing

1350

360-feedback and competencies assessment tools, both individually and

CE

averaging the value obtained from each tool. Based on the HR department professionals’ experience, in the social profiles, and therefore in the nodes

AC

of the social-skill networks, we converted the 0-to-10 values obtained from

1355

these tools into a five-point qualitative scale: “very poor” ([0−2)), “poor” ([2 − 4)), “acceptable”, ([4 − 6)),“good” ([6 − 8)) and “excellent” ([8 − 10]). The simulations corroborated that the objective teammate networks based

on the effort deviation were the most reliable (Fig. 4). The subjective teammate networks based on Q1, Q2 and 360-feedback had poorer and equivalent results. 57

ACCEPTED MANUSCRIPT

1360

And when we combined effort deviation and a subjective metric, no matter which of them, results were very similar to the obtained with the effort deviation (94% vs. 93% of acceptance). Then, we decided to test these latter teammate

CR IP T

network in different pilot experiments. In particular, given the institutional mechanisms to keep updated the 360-feedback assessment, we used these data 1365

instead of questionnaires Q1 and Q2. References

Abreu, R., Premraj, R., 2009. How developer communication frequency relates

AN US

to bug introducing changes. Proceedings of the joint international and annual ERCIM workshops on Principles of software evolution (IWPSE) and software evolution (Evol) workshops - IWPSE-Evol 09 .

1370

Acu˜ na, S., Juristo, N., Moreno, A., 2006. Emphasizing human capabilities in software development. IEEE Software 23, 94–101.

M

Acu˜ na, S.T., G´ omez, M., Juristo, N., 2008. Towards understanding the relationship between team climate and software quality – a quasi-experimental study. Empir Software Eng 13, 401–434.

ED

1375

Acu˜ na, S.T., Juristo, N., 2004. Assigning people to roles in software projects.

PT

Softw: Pract. Exper. 34, 675–696. Adams, S.L., Anantatmula, V., 2010. Social and behavioral influences on team

CE

process. Proj Mgmt Jrnl 41, 89–98. 1380

Albert, R., Barab´ asi, A.L., 2002. Statistical mechanics of complex networks.

AC

Rev. Mod. Phys. 74, 47–97.

Andr´e, M., Baldoqu´ın, M.G., Acu˜ na, S.T., 2011. Formal model for assigning

1385

human resources to teams in software projects. Information and Software Technology 53, 259–275.

Bachmann, A., Bird, C., Rahman, F., Devanbu, P., Bernstein, A., 2010. The missing links: Bugs and bug-fix commits, in: Proceedings of the Eighteenth 58

ACCEPTED MANUSCRIPT

ACM SIGSOFT International Symposium on Foundations of Software Engineering, ACM, New York, NY, USA. pp. 97–106. Bagozzi, R.P., Dholakia, U.M., 2006. Open source software user communities: A

CR IP T

study of participation in linux user groups. Management Science 52, 10991115.

1390

Balthazard, P., Potter, R.E., Warren, J., 2004. Expertise, extraversion and group interaction styles as performance indicators in virtual teams. ACM SIGMIS Database 35, 41.

Bandinelli, S., Di Nitto, E., Fuggetta, A., 1996. Supporting cooperation in the spade-1 environment. IEEE Trans. Software Eng. 22, 841–865.

AN US

1395

Barreto, A., Barros, M.d.O., Werner, C.M., 2008. Staffing a software project: A constraint satisfaction and optimization-based approach. Computers & Operations Research 35, 3073–3089.

Basili, V., Shull, F., Lanubile, F., 1999. Building knowledge through families of

M

experiments. IEEE Trans. Software Eng. 25, 456473.

1400

ED

Bettenburg, N., Hassan, A.E., 2012. Studying the impact of social interactions on software quality. Empir Software Eng 18, 375–431. Bhattacharya, P., Iliofotou, M., Neamtiu, I., Faloutsos, M., 2012. Graph-based

PT

analysis and prediction for software evolution. 2012 34th International Conference on Software Engineering (ICSE) .

1405

CE

Bhattacharya, P., Neamtiu, I., 2010. Fine-grained incremental learning and multi-feature tossing graphs to improve bug triaging. 2010 IEEE International

AC

Conference on Software Maintenance .

Bird, C., Gourley, A., Devanbu, P., Gertz, M., Swaminathan, A., 2006. Mining

1410

email social networks. Proceedings of the 2006 international workshop on Mining software repositories - MSR 06 .

59

ACCEPTED MANUSCRIPT

Blaskovich, J.L., 2008. Exploring the effect of distance: An experimental investigation of virtual collaboration, social loafing, and group decisions. Journal of Information Systems 22, 27–46. Cataldo, M., Herbsleb, J.D., Carley, K.M., 2008. Socio-technical congruence.

CR IP T

1415

Proceedings of the Second ACM-IEEE international symposium on Empirical software engineering and measurement - ESEM 08 .

Cataldo, M., Wagstrom, P.A., Herbsleb, J.D., Carley, K.M., 2006. Identification of coordination requirements. Proceedings of the 2006 20th anniversary

AN US

conference on Computer supported cooperative work - CSCW 06 .

1420

Charette, R.N., 2005. Why software fails [software failure]. IEEE Spectrum 42, 42–49.

Chen, G., Donahue, L.M., Klimoski, R.J., 2004. Training undergraduates to

tion 3, 27–40.

1425

M

work in organizational teams. Academy of Management Learning & Educa-

Chen, W.N., Zhang, J., 2013. Ant Colony Optimization for Software Project

ED

Scheduling and Staffing with an Event-Based Scheduler. IEEE Transactions on Software Engineering 39, 1–17.

PT

Crowston, K., Scozzi, B., 2002. Open source software projects as virtual organisations: competency rallying for software development. IEE Proc., Softw.

1430

CE

149, 3.

Damian, D., Helms, R., Kwan, I., Marczak, S., Koelewijn, B., 2013. The role

AC

of domain knowledge and cross-functional communication in socio-technical

1435

coordination. 2013 35th International Conference on Software Engineering (ICSE) .

DeLone, W.H., McLean, E.R., 2003. The delone and mclean model of information systems success: A ten-year update. Journal of Management Information Systems 19, 9–30.

60

ACCEPTED MANUSCRIPT

Drigas, A., Kouremenos, S., Vrettos, S., Vrettaros, J., Kouremenos, D., 2004. An expert system for job matching of the unemployed. Expert Systems with

1440

Applications 26, 217–224.

CR IP T

Ducheneaut, N., 2005. Socialization in an open source software community: A socio-technical analysis. Computer Supported Cooperative Work (CSCW) 14, 323368. 1445

El Emam, K., Koru, A.G., 2008. A replicated survey of it software project failures. IEEE Software 25, 84–90.

AN US

Espinosa, J.A., DeLone, W., Lee, G., 2006. Global boundaries, task processes

and is project success: a field study. Information Technology & People 19, 345–370. 1450

Feldt, R., Angelis, L., Torkar, R., Samuelsson, M., 2010. Links between the personalities, views and attitudes of software engineers. Information and

M

Software Technology 52, 611–624.

Fenton, N.E., Neil, M., 1999. A critique of software defect prediction models.

1455

ED

IEEE Transactions on Software Engineering 25, 675–689. Fitzpatrick, E.L., Askin, R.G., 2005. Forming effective worker teams with multi-

PT

functional skill requirements. Computers & Industrial Engineering 48, 593– 608.

CE

Gorla, N., Lam, Y.W., 2004. Who should work with whom? Communications of the ACM 47, 79–82.

Grindrod, P., Parsons, M.C., Higham, D.J., Estrada, E., 2011. Communicability

AC

1460

across evolving networks. Phys. Rev. E 83.

Hardiman, L., 1997. Personality types and software engineers. Computer 30, 10–10.

61

ACCEPTED MANUSCRIPT

Herbsleb, J., Mockus, A., 2003. An empirical study of speed and communication in globally distributed software development. IEEE Trans. Software Eng. 29,

1465

481494.

CR IP T

Herbsleb, J.D., 2007. Global software engineering: The future of socio-technical coordination. Future of Software Engineering (FOSE 07) .

Holmstrom, H., Conchuir, E., Agerfalk, P., Fitzgerald, B., 2006. Global software

development challenges: A case study on temporal, geographical and socio-

1470

cultural distance. 2006 IEEE International Conference on Global Software

AN US

Engineering (ICGSE06) .

Katz, L., 1953. A new status index derived from sociometric analysis. Psychometrika 18, 39–43. 1475

Lakhanpal, B., 1993. Understanding the factors influencing the performance of software development groups: An exploratory group-level analysis. Informa-

M

tion and Software Technology 35, 468–473.

Lappas, T., Liu, K., Terzi, E., 2009. Finding a team of experts in social net-

ED

works. Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining - KDD 09 .

1480

PT

Li, C.T., Shan, M.K., 2010. Team formation for generalized tasks in expertise social networks. 2010 IEEE Second International Conference on Social

CE

Computing .

Licorish, S.A., MacDonell, S.G., 2014. Understanding the attitudes, knowledge sharing behaviors and task performance of core developers: A longitudinal

AC

1485

study. Information and Software Technology 56, 1578–1596.

Licorish, S.A., MacDonell, S.G., 2015. Communication and personality profiles of global software developers. Information and Software Technology 64, 113– 131.

62

ACCEPTED MANUSCRIPT

1490

Lievens, F., van Dam, K., Anderson, N., 2002. Recent trends and challenges in personnel selection. Personnel Review 31, 580–601. Linberg, K.R., 1999. Software developer perceptions about software project

CR IP T

failure: a case study. Journal of Systems and Software 49, 177–192.

London, M., Beatty, R.W., 1993. 360-degree feedback as a competitive advantage. Hum. Resour. Manage. 32, 353–372.

1495

Luna-Reyes, L.F., Zhang, J., Ramn Gil-Garca, J., Cresswell, A.M., 2005. Information systems development as emergent socio-technical change: a practice

AN US

approach. Eur J Inf Syst 14, 93105.

MacQueen, J.B., 1967. Some methods for classification and analysis of multivariate observations, in: Proc. of the fifth Berkeley Symposium on Mathematical

1500

Statistics and Probability, University of California Press. pp. 281–297. Madey, G., Freeh, V., Tynan, R., 2002. The open source software develop-

1505

ED

Proceedings , 1806–1813.

M

ment phenomenon: An analysis based on social network theory. AMCIS 2002

Meneely, A., Williams, L., 2011. Socio-technical developer networks. Proceeding of the 33rd international conference on Software engineering - ICSE 11 .

PT

Meneely, A., Williams, L., Snipes, W., Osborne, J., 2008. Predicting failures with developer networks and social network analysis. Proceedings of the

CE

16th ACM SIGSOFT International Symposium on Foundations of software engineering - SIGSOFT 08/FSE-16 .

1510

AC

Moreno, J.L., 1934. Who shall survive?: A new approach to the problem of human interrelations. .

Nagappan, N., Murphy, B., Basili, V., 2008. The influence of organizational

1515

structure on software quality. Proceedings of the 13th international conference on Software engineering - ICSE 08 .

63

ACCEPTED MANUSCRIPT

Newman, M.E.J., 2001. Scientific collaboration networks. ii. shortest paths, weighted networks, and centrality. Phys. Rev. E 64. Ngo-The, A., Ruhe, G., 2009. Optimized resource allocation for software release

1520

CR IP T

planning. IEEE Transactions on Software Engineering 35, 109–123.

Oh, H., Labianca, G., Chung, M.H., 2006. A multilevel model of group social capital. Academy of Management Review 31, 569–582.

Olson, G., Olson, J., 2000. Distance matters. Human-Computer Interaction 15,

AN US

139–178.

Otero, L.D., Centeno, G., Ruiz-Torres, A.J., Otero, C.E., 2009. A systematic approach for resource allocation in software projects. Computers & Industrial

1525

Engineering 56, 1333–1339.

Otero, L.D., Otero, C.E., 2012. A fuzzy expert system architecture for capability assessments in skill-based environments. Expert Systems with Applications

Robertson, I.T., Smith, M., 2001. Personnel selection. Journal of Occupational

ED

1530

M

39, 654–662.

and Organizational Psychology 74, 441–472. Sawyer, S., 2001. Effects of intra-group conflict on packaged software develop-

PT

ment team performance. Inform Syst J 11, 155–178.

CE

Sawyer, S., 2004. Software development teams. Commun. ACM 47, 9599. 1535

Schuler, D., Zimmermann, T., 2008.

Mining usage expertise from version

AC

archives. Proceedings of the 2008 international workshop on Mining software repositories - MSR 08 .

Shin, Y., Meneely, A., Williams, L., Osborne, J.A., 2011. Evaluating com-

1540

plexity, code churn, and developer activity metrics as indicators of software vulnerabilities. IEEE Transactions on Software Engineering 37, 772–787.

64

ACCEPTED MANUSCRIPT

e Silva, L.C., Costa, A.P.C.S., 2013. Decision model for allocating human resources in information system projects. International Journal of Project Management 31, 100–108.

CR IP T

Sowe, S.K., Stamelos, I., Angelis, L., 2008. Understanding knowledge sharing activities in free/open source software projects: An empirical study. Journal

1545

of Systems and Software 81, 431–446.

Sun, J., Tang, J., 2011. A survey of models and algorithms for social influence analysis. Social Network Data Analytics , 177214.

AN US

Surian, D., Liu, N., Lo, D., Tong, H., Lim, E.P., Faloutsos, C., 2011. Rec-

ommending people in developers collaboration network. 2011 18th Working

1550

Conference on Reverse Engineering .

Tamburri, D.A., Lago, P., Vliet, H.v., 2013. Organizational social structures for software engineering. ACM Computing Surveys 46, 135.

M

Tsai, H.T., Moskowitz, H., Lee, L.H., 2003. Human resource selection for software development projects using taguchis parameter design. European Jour-

1555

ED

nal of Operational Research 151, 167–180. Tuckman, B.W., Jensen, M.A.C., 1977. Stages of small-group development

PT

revisited. Group & Organization Management 2, 419–427. Turley, R.T., Bieman, J.M., 1995. Competencies of exceptional and nonexceptional software engineers. Journal of Systems and Software 28, 19–38.

CE

1560

Wolf, T., Schroter, A., Damian, D., Nguyen, T., 2009. Predicting build failures

AC

using social network analysis on developer communication. 2009 IEEE 31st International Conference on Software Engineering .

Xu, J., Christley, S., Madey, G., 2006. Application of social network analysis to

1565

the study of open source software. The Economics of Open Source Software Development , 247–269.

65

ACCEPTED MANUSCRIPT

Xu, R., Wunsch, D., I., 2005. Survey of clustering algorithms. Neural Networks, IEEE Transactions on 16, 645–678. Yang, H.L., Tang, J.H., 2004. Team structure and team performance in is

CR IP T

development: a social network perspective. Information & Management 41,

1570

335–349.

Zhu, H., Zhou, M., Seguin, P., 2006. Supporting software development with

roles. IEEE Transactions on Systems, Man, and Cybernetics - Part A: Sys-

AC

CE

PT

ED

M

AN US

tems and Humans 36, 1110–1123.

66

CR IP T

ACCEPTED MANUSCRIPT

AN US

1575

Roberto Latorre received the B.S. degree in computer engineering and the Ph.D. in computer science and telecommunications from Universidad Aut´ onoma de Madrid, Madrid, Spain, in 2000 and 2008, respectively. He has an extensive experience in software industry, where he worked with multiple companies as IT 1580

architect and software developer. Since 2002 he has been a professor at Escuela

M

Politcnica Superior at Universidad Autnoma de Madrid, where he currently is

AC

CE

PT

ED

Profesor Contratado Doctor.

1585

Javier Su´ a]rez is a software engineer and a researcher at Kymatic Soluciones Informticas. He received the I B.S. degree in computer engineering from Universidad Politcnica de Madrid. He is an expert in Speech technologies and software development.

67