Warning system for online market research – Identifying critical situations in online opinion formation

Knowledge-Based Systems 24 (2011) 824–836

Contents lists available at ScienceDirect

Knowledge-Based Systems journal homepage: www.elsevier.com/locate/knosys

Warning system for online market research – Identifying critical situations in online opinion formation Carolin Kaiser ⇑, Sabine Schlick, Freimut Bodendorf Department of Information Systems, University of Erlangen-Nuremberg, Lange Gasse 20, 90403 Nuremberg, Germany

a r t i c l e

i n f o

Article history: Received 6 August 2010 Received in revised form 19 March 2011 Accepted 20 March 2011 Available online 2 April 2011 Keywords: Warning system Online market research Web Opinion mining Social network analysis Neuro-fuzzy system

a b s t r a c t More and more consumers are relying on online opinions when making purchasing decisions. For this reason, companies must have knowledge of the actual standing of their products on the Web. A warning system for online market research is being proposed which allows the identification of critical situations in online opinion formation. When critical situations are detected, warnings are subsequently sent to marketing managers and thus allowing marketers the ability to initiate preventive measures. The warning system operates on a knowledge base which contains product-related success values, online opinions and patterns of social interactions. This knowledge is acquired using methods coming from information extraction, text mining and social network analysis. Based on this knowledge the warning system judges situations accordingly. For this purpose, a neuro-fuzzy approach is chosen which learns linguistic rules from data. These rules are employed to estimate future situations. The warning system is applied to two scenarios and yields good results. An evaluation shows that all components of the warning system outperform alternative methods. Ó 2011 Elsevier B.V. All rights reserved.

1. Introduction The Internet is an important platform for information exchange. Numerous consumers meet online where they debate over their experiences with various products. By interacting with each other they influence one another’s opinions and purchasing decisions. A survey conducted by the Opinion Research Corporation showed that 66% of consumers rely on online opinions when making purchasing decisions [60]. Several other studies also proved that online opinions have a great influence on consumers’ decision making [61,45,23]. Moreover, some researchers found coherence between online reviews and sales volumes [32,34]. Therefore, it is of vital importance that companies know what opinions are exchanged on the Internet. Since negative word of mouth has a higher informative value [53,24] and a greater effect on purchasing intensions [44,5] than positive word of mouth, it is crucial for companies to identify critical situations at an early stage. Situations become critical for a company when negative opinions are on the verge of spreading. The diffusion of negative opinions poses a reputational and financial threat [12]. It can harm the company’s image and future sales volume. Therefore, it is of great

⇑ Corresponding author. Tel.: +49 911 5302 295; fax: +49 911 5302 379. E-mail addresses: [email protected] (C. Kaiser), sabine. [email protected] (S. Schlick), [email protected] (F. Bodendorf). 0950-7051/$ - see front matter Ó 2011 Elsevier B.V. All rights reserved. doi:10.1016/j.knosys.2011.03.004

importance to alert marketing managers when such a critical situation arises so that they are in a position to initiate counteractive marketing measures. Case studies have shown that late reactions to negative world of mouth can cause considerable damage to a company’s reputation [12]. However, judging the situation correctly is a very difficult task since many influencing factors must be weighed in order to decide if a situation is critical. Polarities of exchanged opinions must be considered, interaction patterns determining the spread of opinions should be observed and opinion leaders as well as the structure of the network also play a major role in opinion diffusion. Success factors concerning the product in question such as the current sales volume should be regarded as well. A comparison with competing products improves the judgment even further. Due to the complex interdependencies of these factors, considerable experience in the field of online market research is required to estimate situations correctly and automating this task poses a great challenge to researchers. The warning system should not only classify situations correctly but should also be able to learn from data and to produce comprehensible results in order to enable an easy usage and a high acceptance. On the basis of these requirements, a neuro-fuzzy approach was implemented for warning marketing managers. The warning system operates on a knowledge base which comprises all influencing factors needed for the classification of the situations. The influencing factors are automatically extracted from the company’s database and the Internet by applying methods coming from

C. Kaiser et al. / Knowledge-Based Systems 24 (2011) 824–836

information extraction, text mining and social network analysis. The warning system learns linguistic rules from the estimation of past situations and applies these rules for judging future situations. Due to their linguistic form, the rules can be easily interpreted by marketing managers. For evaluation purposes, the approach was applied to two scenarios: a reviewing platform where opinions about soccer shoes were exchanged and a social network where experiences with computer games were discussed. The results of these applications are presented in this paper and comparisons to other methods are made to improve validation. The paper is structured as follows: Section 2 introduces related work and describes our contribution. Section 3 gives an overview of the overall approach. The two main components of the approach are presented in Sections 4 and 5. Applications of this approach are illustrated in Section 6. Comparisons to other methods are drawn in Section 7. Finally, Section 8 concludes with a summary and an outlook on future work.

2. Related work Opinion mining aims at discovering attitudes in texts and is applied increasingly to the Internet to reveal consumer opinions. Numerous text mining approaches have been introduced for identifying opinions towards products and their features [26,15,43,46,20,31,33]. However, these approaches only survey opinions at a certain point in time. There are also many papers which monitor the dynamic evolution of online contents. For example, Viermetz et al. [56] propose a method for tracking short-term and long-term trends over time. Tong and Yager [54] outline a system which automatically summarizes online discussions. The summaries contain a linguistic description of the temporal development of contents exchanged in online forums. Bun and Ishizuka [6] introduce a system for tracking emerging topics with information agents employed to detect changes in topics and generate summaries. Zeng et al. [66] describe a system for analyzing user activities on interactive websites. In their experiments they find coherence between user activities and user interests. Huang et al. [25] present an approach for detecting and tracking evolutionary clusters in online communities. New clusters may emerge and old clusters may disappear due to the changing interests of the community’s members. Choudhury et al. [10] analyze the development of online communities with the aid of key groups by identifying groups which are representative for the whole community and infer the dynamic behavior of the community from the behavior of the groups. These papers deal with past developments, but neglect future developments. Some researchers observe online activities in order to predict future sales volumes. For example, Gruhl et al. [21] find correlations between the mentioning of books in blogs and Amazon’s sales ranking for these books. They have developed an algorithm which allows the prognosis of peaks in sales on the basis of how often the books were mentioned in the blogs. Dhar and Chang [16] forecast music sales based on the mentioning of songs and links to musicians. Onishi and Manchanda [41] show that user generated contents are a good indicator for predicting the sale of green tea, movies and cellular phones. All of these approaches again focus on the consequences of online activities but do not predict future behavior. There is also some research on predicting behavior in online communities. For example, Choudhury [9] analyzes past activities of online groups in order to predict future activities. Dastani et al. [13] predict user preferences based on their e-commerce activities. Choudhury et al. [8] propose a method for forecasting the flow of communication in online communities. Kaiser and

825

Bodendorf [27] identify opinion leaders and trends with the aid of communication relationships. As a result, future opinion formation can be estimated. However, only single patterns of user behavior are considered. The existing approaches from the fields of online market research are only appropriate for warning purposes to a limited degree. The monitoring approaches, which keep track of the recent opinion, only detect critical situations when negative consequences have already occurred. The outlined predictive approaches only focus on single aspects of opinion formation but do not judge the overall situation. This approach, however, aims at detecting critical situations at an early stage by taking all relevant factors of opinion formation into account. When studying other disciplines such as medicine or meteorology, approaches were found in which the detection of critical situations at an early stage is possible. These approaches rely on methods coming from soft computing to ensure a timely warning. Murtha [35] for example, uses fuzzy-logic to predict dense fog. The author states that the method could also be used for other scenarios such as the prediction of a snowstorm. Becker et al. [2] also adopt fuzzy-logic. They present a warning system for cardiovascular anesthesia. Other authors employ artificial neural networks for creating a warning system. Boese [7] for example, presents an early warning system for earthquakes. Hamilton and Hufnagel [22] describe a warning system for epileptic fits. Moreover, Fawcett and Provost [18] demonstrate a warning system for detecting telephone fraud. Yang et al. [62] propose a warning system for loan risk assessment. There are also warning systems which are based on neuro-fuzzy systems. Xu and He [63] for example, illustrate a fire alarm system for high-rise buildings. Paetz and Arlt [42] also apply a neuro-fuzzy approach and outline an alarm system for septic shock patients. The presented approach adopts methods from soft computing to the given problem. With the aid of a neuro-fuzzy system, situations are judged as a whole by taking all relevant factors of opinion formation into account. If critical situations are expected to arise, warnings are sent to marketing managers. Thus they are in a position to initiate counteracting measures.

3. Approach The objective of the warning system is to alert marketing managers in critical situations. Situations are considered as critical if negative opinions towards a product are about to spread and to harm the company’s image and sales volume. The warning system operates on a knowledge base which must be fed with knowledge (see Fig. 1). The component for knowledge acquisition collects data from internal and external sources and transforms it into valuable knowledge. Key values characterizing the success of the product in question are extracted from the company’s database. In addition, consumers’ opinions on the product are gathered from the Web and classified as positive, negative or neutral by means of text mining. Furthermore, relationships in online social networks are examined in order to identify influential persons and to determine the network structure. The acquired knowledge is put into a structured form and saved in the knowledge base. The warning system takes this knowledge as input. Situations are judged on the basis of success factors, Web opinions and network characteristics concerning the product. A fuzzy perceptron is employed to learn the rules which allow the differentiation between critical and non-critical situations. Rules learned from past situations are applied for evaluating future situations. If critical situations are detected, warnings are sent to marketing managers. Consequently, marketing actions can be taken in time in order to prevent the spread of negative opinions.

826

C. Kaiser et al. / Knowledge-Based Systems 24 (2011) 824–836

4. Knowledge acquisition 4.1. Information extraction The aim is to extract and integrate the desired success factors for the product in question from multiple source systems. This task may be very time-consuming and complex depending on the source systems involved [14]. Relevant data is stored in different operational systems from different organizations, such as online shops, own stores or independent retailers. These systems can rely on different types of data organization, e.g. flat files or relational databases. Data may be saved in different formats within these systems which makes transformation necessary. For instance, it may be necessary to convert time fields into a standardized format or sales volumes into the same currency. Moreover, an aggregation of these data is needed. For example, the sales volume of different shops must be aggregated to gain the overall sales volume. This process of data extraction and integration must be executed periodically for updating purposes.

4.2. Opinion mining

Fig. 1. Architecture of warning system.

The approach was applied to two scenarios: the reviewing platform ‘‘fussball-forum.de’’ and the social network ‘‘gamestar.de’’. At ‘‘fussball-forum.de’’ opinions about soccer shoes are exchanged. A warning system was created for soccer shoes by adidas and Nike. Opinions on these shoes were collected from ‘‘fussball-forum.de’’ and the sales volumes of these shoes were extracted from the companies’ databases. On the basis of the opinions and the sales volumes of these shoes, a neuro-fuzzy system judges the situations for adidas and Nike as critical or noncritical. At ‘‘gamestar.de’’, members debate over their experiences with computer games. A warning system was implemented for the three games ‘‘Fallout 3’’, ‘‘Deadspace’’ and ‘‘Far Cry 2’’. Opinions on these three games were extracted from ‘‘gamestar.de’’ and the underlying social network of the members of ‘‘gamestar.de’’ was analyzed. A neuro-fuzzy system evaluates the situations for all three games as critical or non-critical by taking the opinions, the influence of the opinion leaders and the structure of the network into account. While under the first scenario knowledge is acquired by information extraction and opinion mining, under the second scenario knowledge is acquired by opinion mining and social network analysis. Under both scenarios, this knowledge is used by a neuro-fuzzy system to assess the situations. If situations are judged critical, warnings are sent to marketing managers. The presented approach builds on previous work by Kaiser et al. [28], however, it exceeds this work in two aspects. In the first place, the knowledge base is further enriched by acquiring data from different data sources. Data from Web 2.0 is combined with knowledge from company databases. Thus a broader spectrum of applications becomes possible and situations can be judged better. In the second place, the methods from the components opinion mining, social network analysis and neuro-fuzzy system are compared to alternative methods. As a result thereof, the warning system is validated in a broader manner.

The goal of opinion mining is to recognize the polarity of the users’ attitudes based on their product evaluations typed in reviews or postings. Three classes of polarities are differentiated: positive, negative and neutral. Text mining techniques are employed to automate the recognition of opinions. The recognition process consists of two succeeding steps [27]. In the first step, features are extracted from the evaluations. In the second step, the polarity of the evaluations is identified on the basis of these features. The feature extraction is based on the linguistic and statistical analysis of the text. First, the text is decomposed into words. After removing the stop words, all remaining words are reduced to their word stem. Those word stems which are especially characteristic of the polarity classes are chosen as features. For this reason, the relative frequency of all word stems is calculated. Word stems which appear frequently in one polarity class but rarely in the other two classes are employed as features. With the aid of these features, the evaluations are classified as positive, negative or neutral. In general, several learning algorithms can be applied for classification such as naïve bayes or decision trees. Support vector machines [11] are used for classifying the users’ evaluations since they enable the processing of many features and yielded successful results in similar projects (e.g. [43,65]. As support vector machines belong to the class of supervised learning algorithms, they require a training data set for learning. A data set consisting of the evaluations with their features and manually assigned polarity classes must be provided. With the aid of this training data set, the parameters of the binary classification functions are learned. In order to differentiate between three classes, three classification functions are learned which separate positive postings from non-positive postings, negative postings from non-negative postings and neutral postings from non-neutral postings. Finally, evaluations are assigned to the class which has the highest probability. In the simplified two-dimensional case, the classification function can be visualized as a line (see Fig. 2). The line divides the evaluations into two classes. Evaluations which lie on the same side of the line belong to the same class. After classification, the opinions of the evaluations are matched to the users. For each user, the average opinion is calculated over a certain period of time. The average opinion ranges from 1 to 1. While a value of 1 indicates that a user only expressed negative opinions towards a product, a value of 1 indicates that a user only made positive remarks about a product.

C. Kaiser et al. / Knowledge-Based Systems 24 (2011) 824–836

827

network to the maximum number of relationships that are possible. The denser a network is, the more information can be exchanged between network members and the quicker opinions can spread throughout the network [3]. 5. Early warning system 5.1. Requirements Fig. 2. Classification of evaluations.

4.3. Social network analysis According to diffusion theory [49], the social network has a great impact on people’s attitudes. People do not form their opinions independently from one another but interactively with other network members. Two factors play a key role in opinion formation: the opinion leaders and the overall network structure. Opinion leaders are central persons in a network who have great influence on the opinions of other network members [29,55]. According to Keller and Berry [30] a large number of network members can be influenced by only a few opinion leaders. Opinion leaders can differ in the degree to which they affect the opinions of others [49]. Centrality indices from social network analysis [59,52] enable the identification of opinion leaders and the determination of their degree of influence. The centrality indices take values ranging from zero to one. A value of one stands for maximum influence, whereas a value of zero stands for minimum influence. Degree centrality indicates how prominent a person is within his neighborhood. Degree centrality is calculated by dividing the number of relationships which a user has by all relationships in the network. Users with high degree centrality have many direct contacts with other network members and are in a position to affect their opinions. Consequently, they are considered as local opinion leaders. In contrast to degree centrality, closeness centrality shows how prominent a person is within the overall network. Closeness centrality of a user is defined as the inverse sum of the distances from this person to all other persons in the network. Users with high closeness centrality are very close to all persons in the network and are able to influence the opinions of many others. Their special position in the network makes them global opinion leaders. Betweenness centrality describes to what degree a person can influence the interactions of other network members. It is calculated as the ratio of the number of shortest paths on which a user lies to the number of all shortest paths. Users with high betweenness centrality lie on many paths between network members and have the possibility of influencing their information exchange. These users are considered as intermediaries. Not only central persons but also the overall structure of the social network has an impact on opinion formation. Centralization and density are important indices from social network analysis which characterize the network structure [59,52]. Centralization indicates how tight the network is around its most central users. The calculation is based on the sum of the differences in centrality of all users from the most central user. In a tightly centralized network, there is a high probability that the opinion of the opinion leaders will spread. There are only a few opinion leaders in the center of the network and many other network members on the periphery. Opinions can spread easily from the leaders to the others [3]. Density characterizes the cohesion within a network. It is calculated as the fraction of the actual number of relationships within a

Before developing an early warning system, the requirements must be defined. Two types of requirements can be distinguished: general requirements which apply to all warning systems and special requirements which apply only to this specific warning system. Therefore general requirements derive from literature. Special requirements result from this particular task. 5.1.1. General requirements 5.1.1.1. Alarm sensitivity. The system must identify critical situations. Otherwise, costs will arise as a consequence of non-identification [19]. In the present case, these costs are losses in sales volume and image. Moreover, the warning must be accurately timed since even a delayed warning would lead to costs. Therefore, the warning must be given as soon as possible. The alarm sensitivity is defined as the ratio of discovered critical situations to all critical situations. 5.1.1.2. Low false alarm rate. An alarm should only be generated in case of real danger. Even false alarms cause costs since managers take counteractive measures [19]. Moreover, managers become desensitized and may not react to a real alarm in future. The false alarm rate is calculated as the ratio of false alarms to all alarms. 5.1.1.3. Trade-off between alarm sensitivity and false alarm rate. Alarm sensitivity and false alarm rate are conflictive [63]. Alarm sensitivity specifies at what point a situation is judged as critical. If the alarm sensitivity is high then the manager is alerted early and an accurately timed warning is given. However, there may be many false alarms. If the false alarm rate is low, the manager is only alerted in case of real danger. However, the warning may not be accurately timed. The balance of alarm sensitivity and false alarm rate must be determined. 5.1.2. Special requirements 5.1.2.1. Learning ability. There are a lot of influencing variables (overall opinion, opinion of the opinion leader, etc.). All these variables must be taken into consideration in order to identify critical situations. These variables differ in different industries and have complex interdependencies. This makes it difficult for experts to express their knowledge in a structured form. Hence the warning system should be capable of learning classification from training data. 5.1.2.2. Using a priori expert knowledge. Due to their professional experience, marketing managers have expert knowledge. This knowledge is often available in the form of linguistic ‘‘if-then rules’’. The system should be able to utilize this knowledge for identifying critical situations. 5.1.2.3. Interpretability. Managers should be in a position to understand the classification of the system. Therefore, the system should not be a black box. On the contrary, it should offer an easily comprehensible explanation of when and why a situation is classified as critical.

828

C. Kaiser et al. / Knowledge-Based Systems 24 (2011) 824–836

5.2. Selection of method The special requirements, i.e. learning ability, usage of a priori expert knowledge and interpretability of classification results must be taken into account when selecting a method for the warning

Fig. 3. Structure of an artificial neural network.

Fig. 4. Fuzzy sets.

system. The discipline of soft computing offers two methods which could be employed: artificial neural networks and fuzzy systems. While artificial neural networks are able to learn classification from data, fuzzy systems enable experts to generate linguistic rules which can be easily interpreted. The model for artificial neural networks is derived from the human brain [4]. The psychologist Rosenblatt [50,51] found that the human brain is capable of associating stimuli with specific responses. On the basis of his findings, he developed the first perceptrons. Perceptrons are three-layered artificial neural networks that are able to calculate classifications by associating input patterns with classes of output patterns [47]. Fig. 3 shows a schematic illustration of such an artificial neural network with its nodes (neurons) and edges (relationships between the neurons). The first layer is called input layer, the second layer is called hidden layer and the third layer is named output layer (one neuron for each class). The impulses are given to the input layer and then pass through the network until they reach the output layer. On the basis of training data, the artificial neural network can learn classification by adjusting network parameters. Although artificial neural networks meet the requirement of being able to learn classifications from data, they are a black box. It is not possible to understand which rules underlie the classification [37]. Fuzzy systems are comprehensible and can use a priori expert knowledge. The theory of fuzzy sets can be traced back to Lotfi Zadeh who published his paper ‘‘Fuzzy Sets’’ in 1965 [64]. In binary logic, an object either belongs to a set or not. In fuzzy logic, on the contrary, an object can belong to several sets with certain degrees of membership (see Fig. 4). Therefore linguistic terms such as ‘‘slightly negative’’ can be modeled. Fuzzy logic enables the formulation of linguistic ‘‘if-then rules’’, e.g. ‘‘if the opinion is slightly negative, then warning’’. Experts can formulate these linguistic rules and the fuzzy system can apply these rules for classification. However, fuzzy systems are not able to learn these rules from data. The warning system must be interpretable, able to utilize a priori expert knowledge and able to learn classification rules from data. Hence a neuro-fuzzy approach is chosen in this work. Neuro-fuzzy systems combine the advantages and minimize the disadvantages of neural networks and fuzzy systems [37]. They are interpretable and can learn linguistic rules for classification from data. Many different neuro-fuzzy models were developed in

Fig. 5. Structure of a fuzzy perceptron.

C. Kaiser et al. / Knowledge-Based Systems 24 (2011) 824–836

the past. For this application the NEFCLASS model (NEuro Fuzzy CLASSification) was selected. This system can learn fuzzy sets and fuzzy rules. In addition, it is also capable of dealing with manually defined rules (a priori expert knowledge) and optimizing them [37].

5.3. Neuro-fuzzy system The NEFCLASS model is a three-layered fuzzy perceptron [38,39]. The structure of the perceptron is illustrated in detail in Fig. 5. The input variables for the classification are represented by the input layer, the fuzzy rules are represented by the hidden layer and the classes are represented by the output layer. In this case, the classes are ‘‘critical situation’’ and ‘‘non-critical situation’’. The weights between the input and the hidden layer represent the linguistic terms [39]. The illustrated fuzzy perceptron contains four fuzzy rules. Rules 1 and 2 classify situations as critical, rules 3 and 4 apply to noncritical situations. According to rule 1 situations are critical if the values of the input variables ‘‘overall opinion’’ and ‘‘opinion of opinion leader’’ are negative (see Fig. 6). The algorithm used by the perceptron for learning the fuzzy rules is an enlargement of the algorithm of Wang and Mendel [4,57,58]. Overlapping hyperboxes construct the feature space and represent the fuzzy sets [37]. Each hyperbox is the n-dimensional Cartesian product of n fuzzy sets [40]. A feature space

If overall opinion is negative and the opinion of the opinion leader is negative then the situation is critical

Fig. 6. Exemplary fuzzy rule.

829

constructed by overlapping hyperboxes is illustrated in Fig. 7. The two diagramed variables have three fuzzy sets: opinion leadership (with the fuzzy sets low, medium, high) and opinion (with the fuzzy sets negative, neutral, positive). The algorithm runs through the training data set twice for learning the fuzzy rules. All possible antecedents (if parts) of the rules are generated in the first run. The arrangement of fuzzy sets which achieves the highest degree of membership is selected for each pattern of the training data set. Then the best consequents (then parts) of the rules are selected in the second run. With the aid of the rules, the input pattern can be classified as critical or non-critical. However, some patterns may still not be classified (triangle) or may be misclassified (circle) (see Fig. 7, left side). In order to improve the classifier, the shape and the position of the fuzzy sets are adapted [39]. After modification of the classifier, more patterns are classified correctly and there are less unclassified patterns (see Fig. 7, right side). To improve the classifier once more, the rule base is pruned by deleting variables or whole rules. By doing so, the rule base is clearer, easier to interpret and can be applied to a broader variety of cases [36]. 6. Application 6.1. Overview The approach was applied to two cases: reviewing platforms and social networks. In the first case, the users do not communicate with one another directly, but simply write and read product evaluations. In the second case, the users interact with each other by submitting postings to forums and giving responses to the postings of other users. Depending on the different types of communication platforms, different variables must be considered when judging situations as critical or non-critical. The following sections outline one application for each case. For the first application, opinions from the reviewing platform and sales volumes from the companies’ databases are considered in order to assess the situations. For the second application, opinions as well as influential persons

Fig. 7. Classification of situations after rule learning (left side) and after fuzzy set modification (right side).

830

C. Kaiser et al. / Knowledge-Based Systems 24 (2011) 824–836

Table 1 Input variables with fuzzy sets. Opinion on Predator compared to previous month

Sales volume of Predator

Opinion on Mercurial

Opinion on Mercurial compared to previous month

Sales volume of Mercurial

Negative Slightly negative Neutral Slightly positive Positive

Decreased Slightly decreased Unchanged Slightly increased Increased

Very low Low Medium High Very high

Negative Slightly negative Neutral Slightly positive Positive

Decreased Slightly decreased Unchanged Slightly increased Increased

Very low Low Medium High Very high

Jun-08

May-08

Apr-08

Mar-08

Feb-08

Jan-08

Nov-07

Dec-07

Oct-07

Sep-07

Aug-07

Jul-07

Jun-07

May-07

Apr-07

Mar-07

Feb-07

Sales Volume (no scale since data is confidential)

0.12 0.11 0.1 0.09 0.08 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0 -0.01 -0.02 Jan-07

Opinion

Opinion on Predator

Time Opinion on Mercurial Sales Volume of Mercurial

Opinion on Predator Sales Volume of Predator

Fig. 8. Development of input variables.

and structural properties of the social network are taken into consideration in order to estimate the situations. For both applications, a neuro-fuzzy system judges the situations as critical or non-critical. If critical situations are detected warnings are sent to the marketing managers. While the warning system for the first application comprises the components information extraction, opinion mining and neuro-fuzzy system, the warning system for the second application relies on the components opinion mining, social network analysis and neuro-fuzzy system. 6.2. Reviewing platform 6.2.1. Data collection The data sets were extracted from the reviewing platform ‘‘fussball-forum.de’’, where opinions on soccer shoes are exchanged. 407 postings consisting of 2095 sentences were extracted from January 2007 to June 2008. The users and their opinions towards the soccer shoes ‘‘Predator’’ from adidas and ‘‘Mercurial’’ from Nike were determined. The opinions mentioned in the sentences were classified as positive, negative or neutral by means of text mining with an average accuracy of 73%. Afterwards, the overall opinion was calculated for each month. Furthermore, the sales volume of the two shoes was also taken from the companies’ databases and loaded into the knowledge base. 6.2.2. Training set As the neuro-fuzzy model learns on a supervised basis, all data sets must be classified manually. For each point in time the marketing manager must decide whether the situation is critical or not. In the case of reviewing platforms, the sales volume and the

overall opinion as well as the overall opinion in comparison to the previous month are taken into consideration for both one’s own product as well as for the competitor’s product. For each variable, five fuzzy sets are chosen. Table 1 shows the six variables with their fuzzy sets. The development of these variables must be taken into account when judging situations as critical or non-critical. Fig. 8 shows the time series of these variables. For example, the managers of adidas should be warned in June 2007 that the situation is critical because the opinion and the sales volume of their own product (‘‘Predator’’) have decreased in comparison to the previous month. In contrast, these variables remain unchanged for the competitor’s product (‘‘Mercurial’’). In March 2008 the situation is not critical to adidas. Although the sales volume of the competitor’s product (‘‘Mercurial’’) is high, it does not have a negative effect on adidas since the sales volume of their own product (‘‘Predator’’) is also high. Moreover, the opinion about their own product (‘‘Predator’’) has increased compared to the previous month, whereas the opinion about the competitor’s product (‘‘Mercurial’’) has decreased in comparison with the previous month.

6.2.3. Rules Based on the training data sets, rules for both cases (warning for adidas or warning for Nike) are learned by the neuro-fuzzy system. For both of them, the rule base is easily interpretable and clear with each rule base consisting of five rules. Fig. 9 demonstrates an extraction of the rule base for the shoe ‘‘Predator’’. Rule 1 specifies a situation which is critical to adidas. The managers must be alerted since the opinion towards ‘‘Predator’’ (own product) has decreased slightly in comparison to the previous month. Moreover,

831

C. Kaiser et al. / Knowledge-Based Systems 24 (2011) 824–836

Rule 1: If opinion on Predator compared to previous month has slightly decreased and sales volume of Predator is very low and sales volume of Mercurial is low then a warning is issued. Rule 2:

The aim of this application was to test whether the installation of an early warning system for adidas and Nike is beneficial. Therefore, only a few data sets were considered in this first step. The results revealed that it is advantageous to install such a system and henceforth more data sets will be employed in future.

If opinion on Predator is slightly positive and opinion on Predator compared to previous month has slightly increased and sales volume of Predator is very low and sales volume of Mercurial is medium then no warning is issued.

Fig. 9. Rules for adidas.

the sales volume of ‘‘Predator’’ is very low, while the sales volume of ‘‘Mercurial’’ (competing product) is low. Rule 2 describes situations which are not critical to adidas even though the sales volume of ‘‘Mercurial’’ (medium) is higher than the sales volume of ‘‘Predator’’ (very low). However, the opinion towards ‘‘Predator’’ is slightly positive and has also increased slightly in comparison to the previous month. Therefore there is no need to alert adidas’ managers, since the sales volume of ‘‘Predator’’ is expected to increase in the future.

6.2.4. Results The classifiers for adidas’ ‘‘Predator’’ and Nike’s ‘‘Mercurial’’ worked well (see Table 2). The classifier for the shoe ‘‘Predator’’ classified 80% of the situations correctly and the one for the shoe ‘‘Mercurial’’ classified 70% correctly (both cross-validated 10 times). The false alarm rate for the shoe ‘‘Predator’’ was 0% and the alarm sensitivity was 77.78%. This means that the managers were warned in 77.78% of all critical situations putting them in a position to intervene. There were no alarms in non-critical situations. For the shoe ‘‘Mercurial’’ the false alarm rate was 40% and the alarm sensitivity was 100%. Hence the managers were alerted in 40% of the non-critical situations. However, they were warned in all situations that were critical for their company. The results for the shoe ‘‘Predator’’ are good and the results for the shoe ‘‘Mercurial’’ are acceptable as well. However, it should be carefully considered whether to trade off the high alarm sensitivity for the reduction of the false alarm rate. Fundamentally, it is a matter of weighing the costs of false alarms against the costs resulting from undetected critical situations.

Table 2 Results for reviewing platform. Shoe

Data sets

Correctly classified (%)

Alarm sensitivity (%)

False alarm rate (%)

Predator Mercurial

18 18

80.0 70.0

77.8 100.0

0.0 40.0

6.3. Social network 6.3.1. Data collection The data for the second case was collected from the social network ‘‘gamestar.de’’, where members discuss their experiences on computer games. The observation period was from the fifth of October to the 28th of November 2008. During this time, 3776 postings concerning the game ‘‘Fallout 3’’, 1350 postings about the game ‘‘Dead Space’’ and 1470 postings about the game ‘‘Far Cry 2’’ were extracted. For each game, a series of daily networks were constructed by linking those users who had posted their opinions directly before or after one another on the same day. The opinions were classified as positive, negative or neutral with the aid of text mining methods. An average accuracy of 77% was achieved for opinion classification [28]. Moreover, a social network analysis was executed in order to determine the influence of the opinion leaders and the structure of the network. 6.3.2. Training set In order to judge situations in social networks, the opinions of its members and the characteristics of the network must be taken into account. The opinions of the opinion leaders (local opinion leader, global opinion leader and intermediary), the influence of the opinion leaders, the likelihood that the opinion of the opinion leader will diffuse, speed of diffusion and the overall opinion of the network are considered. For each variable, three fuzzy sets were chosen (see Table 3). Figs. 10 and 11 show two points in time from the social network concerning the game ‘‘Dead Space’’. Fig. 10 illustrates a situation that is judged as non-critical. The opinion of the global opinion leader (user 26) is positive and his influence is high (closeness centrality 0.65). The opinion of user 25, who is the local opinion leader and the intermediary, is also positive. His influence is medium (degree centrality 0.57, betweenness centrality 0.48) and the overall opinion is neutral to slightly positive (0.24). The likelihood that the opinion of the opinion leader will diffuse is medium (centralization 0.4) and the speed of diffusion is low (density 0.18). The future overall opinion within the network is expected to remain positive. In contrast, the situation illustrated in Fig. 11 is judged as critical. The opinion of the local opinion leader and intermediary (user 8) as well as the opinion of the global opinion leader (user 9) is neutral. The influence of user 8 is high (degree centrality 0.75) to medium (betweenness centrality 0.41). The influence of

Table 3 Input variables with fuzzy sets. Influence of local opinion leader

Opinion of local opinion leader

Influence of global opinion leader

Opinion of global opinion leader

Influence of intermediary

Opinion of intermediary

Low Medium High

Negative Neutral Positive

Low Medium High

Negative Neutral Positive

Low Medium High

Negative Neutral Positive

Speed of diffusion

Likelihood that opinion of opinion leader will diffuse

Overall opinion

Low Medium High

Low Medium High

Negative Neutral Positive

832

C. Kaiser et al. / Knowledge-Based Systems 24 (2011) 824–836

Rule 1: If the opinion of global opinion leader is negative and the opinion of local opinion leader is negative and the influence of local opinion leader is high and likelihood that opinion of the opinion leader will diffuse is medium and overall opinion is neutral then a warning is issued. Rule 2: If opinion of the global opinion leader is positive and likelihood that opinion of the opinion leader will diffuse is low and overall opinion is neutral then no warning is issued. Fig. 12. Rules for the game ‘‘Fallout 3’’.

Fig. 10. Non-critical situation in social network.

user 9 is high (closeness centrality 0.57) and the likelihood that the opinion of the opinion leader will diffuse is high (centralization 0.47). The speed of diffusion is medium (density 0.36) and the overall opinion is neutral to slightly negative ( 0.11). Since neutral opinions can be a sign of disinterest, the situation is evaluated as critical.

6.3.3. Rules For each of the three games a clear and easily interpretable rule base was learned by the neuro-fuzzy system. The rule base for the game ‘‘Dead Space’’ consists of eight rules, the one for the game ‘‘Fallout 3’’ comprises nine rules and the one for the game ‘‘Far Cry 2’’comprises three rules. Fig. 12 shows two rules of the rule base for the game ‘‘Fallout 3’’. Rule 1 states that the managers must be warned, if the opinions of the global leader and the local leader are negative and the influence of the local opinion leader is high. Moreover, the likelihood that the opinion of the

opinion leader will diffuse must be medium and the overall opinion must be neutral. In these situations there is the danger that the negative opinion of the opinion leaders will diffuse and that the already neutral overall opinion will become worse, i.e. negative. Rule 2 describes a situation in which the overall opinion is neutral. In contrast to the previous rule, the opinion of the global opinion leader is positive. Even though the likelihood that the opinion of the opinion leader will diffuse is low, it is not necessary to alert the manager. In such cases there is no danger of the overall opinion becoming negative. 6.3.4. Results The classifiers for all three games worked very well. Table 4 shows the classification results for each game with the best results being achieved by ‘‘Dead Space’’. 90.3% of all situations were classified correctly and 100% of the critical situations were identified. This means that in all critical situations a warning was generated and that the managers were alerted before the overall opinion became too negative and sales went down. A false alarm was only generated in 11.8% of the non-critical situations. The classifier for the game ‘‘Fallout 3’’ yielded very good results as well. It classified 93.5% of all situations correctly, identified 88.9% of the critical situations and only generated 5.9% false alarms. This means that the managers were warned in 88.9% of all critical situations and only in 5.9% of non-critical situations. For the game ‘‘Far Cry 2’’, the classification results were less successful. Due to the small training set, only 70.6% of all situations were classified correctly. However, in 88.2% of the critical situations the managers were alerted and the false alarm rate of 25.0% was also acceptable. 7. Comparison to other methods 7.1. Opinion mining

Fig. 11. Critical situation in social network.

In order to evaluate the opinion mining component, support vector machines were compared to naïve bayes and decision tree C4.5 [48]. Both methods had often been applied successfully in the fields of text classification [17,1]. Naïve bayes is a probabilistic classification method. It estimates the conditional probability of a posted opinion on the basis of the features of the text. The opinion is assigned to the polarity class that has the highest probability.

833

C. Kaiser et al. / Knowledge-Based Systems 24 (2011) 824–836

Decision trees learn classification models which resemble the form of trees. The tree is generated by a recursive procedure which partitions the opinions according to their textual features. The branches of the tree determine which textual features lead to which opinion class. An opinion is assigned to a polarity class by following the branches of the tree. Table 5 shows the opinion mining results for support vector machines, naïve bayes and decision trees for the two cases of applications. The depicted accuracies were measured during a five-fold cross validation. For the case studies ‘‘Predator’’ and ‘‘Mercurial’’, support vector machines achieved an accuracy of 73.6%, whereas naïve bayes and decision trees only achieved 65.8%. For the case study ‘‘Dead Space’’, ‘‘Fallout 3’’ and ‘‘Far Cry 2’’ support vector machines yielded an accuracy of 77%, whereas naïve bayes and decision trees only yielded accuracies of 69.7% and 67.2% respectively. Support vector machines outperform the other methods by far.

7.2. Social network analysis The social network analysis was also compared to other methods. First, an expert panel determined the opinion leaders with respect to the number of postings per user, the contents of the postings and the influence of each user. The experts read each posting and then decided unanimously which user the opinion leader of each day is. Afterwards, the social network analysis and two other methods were compared to the human decision. Since opinion leaders have more friends than others [29], the users with the highest number of friends are identified as opinion leaders. This means that from all users who submitted a posting on the day in question, the one with the most friends is considered as the opinion leader.

Table 4 Results for social network. Game

Data sets

Correctly classified (%)

Alarm sensitivity (%)

False alarm rate (%)

Dead space Fallout 3 Far Cry 2

56 45 33

90.3 93.5 70.6

100.0 88.9 88.2

11.8 5.9 25.0

Moreover, opinion leaders have great influence on the opinions of other network members [29,55]. Therefore, the user whose opinion is adopted most frequently is considered as opinion leader. This adoption can be measured on the basis of the posting chains in the threads. This means that the opinion leader is the user whose posting is followed by the highest number of postings sharing his opinion. In all cases the social network analysis works best. Table 6 shows the results of the evaluation. For the game ‘‘Dead Space’’ 84.3% of the opinion leaders can be identified, for the game ‘‘Fallout 3’’ 77.8% and for the game ‘‘Far Cry 2’’ 78.8%. 7.3. Neuro-fuzzy system In order to improve validation, the neuro-fuzzy approach was compared to a fuzzy system and artificial neural network. A panel of three experts was consulted to model the fuzzy expert rules. In an iterative process of several days they developed a rule base for each case of application. The experts succeeded in working out a clear rule base for all cases. The rule base for the shoe ‘‘Predator’’ consists of five rules and the rule base for the shoe ‘‘Mercurial’’ comprises four rules. The rule bases for the games ‘‘Dead Space’’, ‘‘Fallout 3’’ and ‘‘Far Cry 2’’ contain three, four and three rules, respectively. The rule base can be easily interpreted by marketing managers since it is based on human expert knowledge. All in all, the rule bases of the neuro-fuzzy model are larger than, or as large as, the ones of the fuzzy system (see Table 7). However, the rule bases of both models are clear and easily interpretable. All five fuzzy classifiers work well. The classifiers for the shoes ‘‘Predator’’ and ‘‘Mercurial’’ classified 66.7% of the situations correctly and the classifiers for the games ‘‘Dead Space’’, ‘‘Fallout 3’’ and ‘‘Far Cry 2’’ classified 84.3%, 82.2% and 72.2% correctly (see Table 8). However, the classifiers of the neuro-fuzzy model work better in all cases except for one (‘‘Far Cry 2’’). On average the false alarm rate of the fuzzy system is similar to that of the neuro-fuzzy system (see Appendix). However, the neuro-fuzzy system has a higher alarm sensitivity and it surpasses the fuzzy system in this respect. At large the neuro-fuzzy system outplays the fuzzy model since it is easier and less time-consuming to create and produces better classification results.

Table 7 Number of rules.

Table 5 Opinion mining results. Case of application

Support vector machines (%)

Naïve bayes (%)

Decision tree (%)

Predator, Mercurial Dead space, Fallout 3, Far Cry 2

73.6 77.0

65.8 69.7

65.8 67.2

Predator Mercurial Dead space Fallout 3 Far Cry 2

Neuro-fuzzy system

Fuzzy system

5 5 8 9 3

5 4 3 4 3

Table 8 Comparison of classification results. Table 6 Identification of opinion leaders. Case of application

Social network analysis (%)

Friends (%)

Posting chains (%)

Dead space Fallout 3 Far Cry 2

84.3 77.8 78.8

70.6 48.9 60.6

68.6 55.6 51.5

Case of application

Data sets

Correctly classified (10-fold cross validation) Neuro-fuzzy system (%)

Fuzzy system (%)

Artificial neural network (%)

Predator Mercurial Dead Space Fallout 3 Far Cry 2

18 18 56 45 33

80.0 70.0 90.3 93.5 70.6

66.7 66.7 84.3 82.2 72.7

72.2 72.2 86.2 86.7 63.6

834

C. Kaiser et al. / Knowledge-Based Systems 24 (2011) 824–836

In addition, the neuro-fuzzy system was also compared to an artificial neural network. For this purpose, a three-layered perceptron was trained for each case of application. In contrast to the neuro-fuzzy model and the fuzzy model, the artificial neural network does not construct a rule base. The artificial neural network is a black box and there is no possibility of interpreting the classification results. The classification accuracies achieved by the artificial neural network are depicted in Table 8. The classifier of the artificial neural network for the soccer shoe ‘‘Mercurial’’ classified somewhat better (72.2%) than the one of the neuro-fuzzy model. The classifiers for the shoe ‘‘Predator’’ and the three games also yielded high accuracies ranging from 63.6% to 86.7%. However, these classifiers are less successful than those constructed by the neuro-fuzzy model. The alarm sensitivity of the neural network is lower than that of the neuro-fuzzy system in all cases (see Appendix). Also, the false alarm rate is not better than the one of the neuro-fuzzy system. All in all, the results proved that the neuro-fuzzy system is superior to the artificial neural network. Except for one case (‘‘Mercurial’’), the classification results are better than those of the artificial neural network. Even alarm sensitivity and false alarm rate are better. In addition, the classifier constructed by the neural network is not interpretable. In conclusion it can be said that the neuro-fuzzy model is superior to the fuzzy model and the artificial neural network in all aspects. The fuzzy system comes in second and the artificial neural network comes in third. 8. Conclusion The outlined warning system allows the identification of critical situations for a company’s product and the immediate alerting of marketing managers. Knowledge from the company’s database and online discussions is acquired for judging situations. Product-related success factors are extracted from databases, online opinions are recognized by means of text mining and network behavior is examined with the aid of social network analysis. On the basis of this knowledge, situations are classified as critical or non-critical. A neuro-fuzzy system is chosen for classification. The neuro-fuzzy system learns linguistic rules from past situations which enable the estimation of future situations. The rules can be easily understood by marketing managers. The outlined approach contributes to the expansion of existing work in the field of online market research in two ways. In contrast to existing monitoring systems, which only detect critical situations when they have already arisen, this approach enables a timely warning. Thus marketers are still in a position to initiate preventive measures. In comparison to predictive ap-

proaches, which only focus on single aspects of opinion formation, this approach takes all relevant factors into account. Thus situations are judged as a whole and this enhances the recognition of situations which are really critical for a company. If marketing managers are alerted at a point in time when situations are getting really critical they are in a position to initiate counteracting measures and are able to prevent negative opinions from being spread. Marketing measures could include private or public discussions with the users of the online platform in question. Listening to the users’ complaints and wishes may not only lead to product improvements but also to an enhancement of the company’s image. Consumers might be more satisfied if they get the feeling that their opinions are being taken into consideration. The application of the neuro-fuzzy system to the two scenarios produced very good results. A classification accuracy ranging from 70% to 93% was achieved in a 10-fold cross validation. The alarm sensitivity, ranging from 77% to 100%, shows that most of the critical situations could be detected and that marketing managers could be warned in time. The false alarm rate, ranging from 0% to 40%, indicates that only a few non-critical situations were wrongly judged as critical. In order to improve validation, the methods of all components of the warning system were compared to alternative methods. Support vector machines employed for opinion mining were compared to naïve bayes and decision tree C4.5. For both cases of application, support vector machines outperformed naïve bayes and decision trees. Social network analysis applied for identifying opinion leaders was compared to methods regarding friends and posting chains. The application revealed that social network analysis is the best method. The neuro-fuzzy system used for identifying critical situations was compared to a fuzzy system and an artificial neural network. Higher performance combined with learning ability and interpretability make the neuro-fuzzy system the preferred method. A future task is to expand the application data sets with more data collection in order improve validation. Moreover, the approach will be extended. The warning system will be augmented by a decision support system. In the case of critical situations, marketing managers should not only get warnings but should also be provided with recommendations for appropriate actions. Recommendations could involve Web communications, marketing campaigns or product improvements.

Appendix A Results of neuro-fuzzy system, fuzzy system and artificial neural network

Product

Data Correctly classified Alarm sensitivity False alarm rate sets Neuro-fuzzy Fuzzy Artificial neural Neuro-fuzzy Fuzzy Artificial Neuro-fuzzy Fuzzy Artificial system (%) system (%) network (%) system (%) system (%) neural system (%) system neural network (%) (%) network (%)

Predator Mercurial Dead Space Fallout 3 Far Cry 2

18 18 56 45 33

80.0 70.0 90.3 93.5 70.6

66.7 66.7 83.3 82.2 72.2

72.2 72.2 86.2 86.7 63.6

77.8 100.0 100.0 88.9 88.2

66.7 66.7 86.7 66.7 88.2

66.7 87.5 90.0 83.3 75.0

0.0 40.0 11.8 5.9 25.0

14.3 0.0 10.4 7.7 25.0

25.0 30.0 12.9 16.7 40.0

C. Kaiser et al. / Knowledge-Based Systems 24 (2011) 824–836

References [1] C. Apte, F. Damerau, S.M. Weiss, Text mining with decision trees and decision rules, in: Proceedings of the Conference on Automated Learning and Discovery, Pittsburgh, 1998. [2] K. Becker, G. Rau, H. Kaesmacher, M. Petermeyer, G. Kalff, H.-J. Zimmermann, Fuzzy logic approaches to intelligent alarms, Engineering in Medicine and Biology Magazine, IEEE 13 (5) (1994) 710–716. [3] F. Bodendorf, C. Kaiser, Detecting opinion leaders and trends in online social networks, in: Proceedings of the Second Workshop on Social Web Search and Mining, Hong Kong, 2009. [4] C. Borgelt, F. Klawonn, R. Kruse, D. Nauck, Neuro-fuzzy-systeme: Von den Grundlagen künstlicher Neuronaler Netze zur Kopplung mit Fuzzy Systemen, [engl.: Neuro-Fuzzy-Systems: Foundations of the combination of neural networks and fuzzy-systems.], third ed., Vieweg, Wiesbaden, 2003. [5] J.J. Brown, P.H. Reingen, Social ties and word-of-mouth referral behavior, Journal of Consumer Research 14 (1987) 350–362. [6] K.K. Bun, M. Ishizuka, Emerging topic tracking system in WWW, KnowledgeBased Systems 19 (3) (2006) 164–171. [7] M. Boese, Earthquake Early Warning for Istanbul Using Artificial Neural Networks, Ph.D. Thesis, University of Karlsruhe, Germany, 2006. [8] M.D. Choudhury, H. Sundaram, A. John, D.D. Seligmann, Contextual prediction of communication flow in social networks, in: Proceedings of the IEEE/WIC/ ACM international Conference on Web Intelligence, IEEE Computer Society, Washington, DC, 2007, pp. 57–65. [9] M.D. Choudhury, Modelling and predicting group activity over time in online social media, in: Proceedings of the 20th ACM Conference on Hypertext and Hypermedia, Torino, 2009. [10] M.D. Choudhury, H. Sundaram, A. John, D.D. Seligmann, Which are the representative groups in a community? Extracting and characterizing key groups in blogs. ACM student research competition, HyperText ’09, 2009. [11] C. Cortes, V.N. Vapnik, Support vector networks, Machine Learning 20 (1995) 273–297. [12] W.T. Coombs, Protecting Organization reputations during a crisis: The development and application of situational crisis communication theory, Corporate Reputation Review 10 (2007) 163–176. [13] M. Dastani, N. Jacobs, C.M. Jonker, J. Treur, Modeling user preferences and mediating agents in electronic commerce, Knowledge-Based Systems 18 (7) (2005) 335–352. [14] T. Dasu, T. Johnson, Exploratory Data Mining and Data Cleaning, John Wiley & Sons, Hoboken, 2003. [15] K. Dave, S. Lawrence, D. Pennock, Mining the peanut gallery: opinion extraction and semantic classification of product reviews, in: Proceedings of the 12th International Conference on World Wide Web, 2003. [16] V. Dhar, E. Chang, Does Chatter Matter? The Impact of User-Generated Content on Music Sales, Technical Report, Leonard N. Stern School of Busi-ness, New York University, 2007. [17] P. Domingos, M.J. Pazzani, On the optimality of the simple Bayesian classifier under zero-one loss, Machine Learning 29 (2–3) (1997) 103–130. [18] T. Fawcett, F. Provost, Adaptive fraud detection, Data Mining and Knowledge Discovery 1 (1997) 291–316. [19] T. Fawcett, F. Provost, Activity monitoring: noticing interesting changes in behavior, in: Proceedings of the Fifth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining Table, San Diego, California, United States, 1999, pp. 53–62. [20] N. Glance, M. Hurst, K. Nigam, M. Siegler, R. Stockton, T. Tomokiyo, Deriving marketing intelligence from online discussion, in: Proceedings of the 11th ACM SIGKDD International Conference on Knowledge Discovery in Data Mining, Chicago, 2005, pp. 419–428. [21] D. Gruhl, R. Guha, R. Kumar, J. Novak, A. Tomkins, The predictive power of online chatter, in: Proceedings of the 11th ACM SIGKDD International Conference on Knowledge discovery in Data Mining, Chicago, 2005, pp. 78–87. [22] Hamilton, Hufnagel, Artificial neural networks detect epileptic attacks, in: H.G. Schuster (Ed.), Applications of Neural Networks, VCH Verlagsgesellschaft, Weinheim, 1992, pp. 173–178. [23] D. Houser, J. Wooders, Reputation in auctions: Theory and evidence from eBay, Journal of Economics and Management Strategy 15 (2006) 353–369. [24] P.M. Herr, F.R. Kardes, J. Kim, Effects of word-of-mouth and product-attribute information on persuasion: an accessibility-diagnosticity perspective, Journal of Consumer Research 17 (1991) 454–462. [25] Y. Huang, S. Liu, Y. Wang, Online detecting and tracking of the evolution of user communities, in: Third International Conference on Natural Computation, 2007, pp. 681–685. [26] C. Kaiser, Combining text mining and data mining for gaining valuable knowledge from online reviews, IADIS International Journal on WWW/ Internet 6 (2) (2009) 63–78. [27] C. Kaiser, F. Bodendorf, Opinion and relationship mining in online forums, in: Proceedings of the 2009 IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent Agent Technology, Milan, 2009, pp. 128–131. [28] C. Kaiser, S. Schlick, F. Bodendorf, Discovering critical situations in online social networks – a neuro fuzzy approach to alert marketing managers, in: Proceedings of the International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management, Valencia, 2010. [29] E. Katz, P.F. Lazarsfeld, Personal Influence, The Part Played by People in the Flow of Mass Communication, Glencoe, Free Press, 1955.

835

[30] E.B. Keller, J. Berry, The Influentials, Free Press, New York, 2003. [31] S.-M. Kim, E. Hovy, Crystal: analysing predictive opinions on the Web, in: Proceedings of the 2007 Joint Conference on the Empirical Methods Natural Language Processing and Computational Natural Language Learning, Prague, 2007, pp. 1056–1064. [32] Y. Liu, X. Huang, A. An, X. Yu, ARSA: a sentiment-aware model for predicting sales performance using blogs, in: Proceedings of the 30th Annual International ACM SIGIR Conference, 2007, ACM, Amsterdam. [33] M.P. O’Mahony, B. Smyth, A classification-based review recommender, Knowledge-Based Systems 23 (4) (2010) 323–329. [34] G. Mishne, N. Glance, Predicting movie sales from blogger sentiment, in: Proceedings of the AAAI 2006 Spring Symposium on Computational Approaches to Analyzing Weblogs, 2006. [35] J. Murtha, Applications of fuzzy logic in operational meteorology, in: Scientific Services and Professional Development Newsletter, Canadian Forces Weather Service, 1995, pp. 42–54. [36] U. Nauck, Design and Implementation of a Neuro-Fuzzy Data Analysis Tool in Java, Diploma Thesis, University of Braunschweig, Braunschweig, 1999. [37] D. Nauck, F. Klawonn, R. Kruse, Foundations of Neuro-fuzzy Systems, JohnWiley & Sons, Chichester, 1997. [38] D. Nauck, R. Kruse, A fuzzy perceptron as a generic model for neuro-fuzzy approaches, in: Fuzzy Systeme ’94, 1994. [39] D. Nauck, R. Kruse, NEFCLASS – a neuro-fuzzy approach for the classification of data, in: K.M. George, J.H. Carrol, E. Deaton, D. Oppenheim, J. Hightower (Eds.), Proceedings of the 1995 ACM Symposium on Applied Computing, ACM Press, Nashville, 1995, pp. 26–28. [40] D. Nauck, R. Kruse, A neuro-fuzzy method to learn fuzzy classification rules from data, Fuzzy Sets and Systems 1997 (89) (1997) 277–288. [41] H. Onishi, P. Manchanda, Marketing Activity, Blogging and Sales, Technical Report, Ross School of Business, University of Michigan, 2009. [42] J. Paetz, B. Arlt, A neuro-fuzzy based alarm system for septic shock patients with a comparison to medical scores, in: A. Colosimo, A. Giuliani, P. Sirabella (Eds.), Proceedings of the Third International Symposium of Medical Data Analysis (ISMDA 2002), Rome, Italy, LNCS, vol. 2526, Springer-Verlag, 2002, pp. 42–52 (Preprint-Version). [43] P. Pang, L. Lee, S. Vaithyanathan, Thumbs up? Sentiment classification using machine learning techniques, in: Proceedings of the Conference on Empirical Methods in Natural Language Processing, ACM, 2002, pp. 79–86. [44] C. Park, T.M. Lee, Information direction, website reputation and eWOM effect: a moderating role of product type, Journal of Business Research 62 (1) (2009) 61–67. [45] D.H. Park, J. Lee, I. Han, The effect of on-line consumer reviews on consumer purchasing intention: the moderating role of involvement, International Journal of Electronic Commerce 11 (2007) 125–148. [46] A.-M. Popescu, O. Etzioni, Extracting product features and opinions from reviews, in: Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/ EMNLP), 2005, pp. 339–346. [47] K.L. Priddy, P.E. Keller, Artificial Neural Networks: An Introduction, SPIE Press, Bellingham, 2005. [48] J.R. Quinlan, C4.5. Programs for Machine Learning, Morgan Kaufman, San Mateo, 1992. [49] E. Rogers, Diffusion of Innovations, fifth ed., Free Press, New York, 2003. [50] F. Rosenblatt, The perceptron: a probabilistic model for information storage and organization in the brain, Psychological Review 65 (1958) 386– 408. [51] F. Rosenblatt, Principles of Neurodynamics: Perceptrons and the Theory of Brain Mechanisms, Spartan Books, Washington, DC, 1961. [52] J. Scott, Social Network Analysis – A Handbook, SAGE, London, 2000. [53] S. Sen, D. Lerman, Why are you telling me this? An examination into negative consumer reviews on the web, Journal of Interactive Marketing 21 (4) (2007) 76–94. [54] R.M. Tong, R.R. Yager, Characterizing attitudinal behaviors in on-line opensource, in: Proceedings of Association for the Advancement of Artificial Intelligence, Spring Symposium 2004, Atlanta. [55] T.W. Valente, Network Models of the Diffusion of Innovations, Hampton Press, Cresskill, 1999. [56] M. Viermetz, M. Skubacz, C.-N. Ziegler, D. Seipel, Tracking topic evolution in news environments, in: 10th IEE Conference on E-commerce Technology and the Fifth IEEE Conference on Enterprise Computing, E-Commerce and EServices, 2008, pp. 215–220. [57] L.-X. Wang, J.M. Mendel, Generating rules by learning from examples, in: International Symposium on Intelligent Control, IEEE Press, Piscataway, NJ, USA, 1991, pp. 263–268. [58] L.-X. Wang, J.M. Mendel, Generating fuzzy-rules by learning from examples, IEEE Transactions on Systems, Man, and Cybernetics 22 (6) (1992) 1414– 1427. [59] S. Wassermann, K. Faust, Social Network Analysis – Methods and Applications, Cambridge University Press, Cambridge, 1999. [60] C. Werbler, C. Harris, Online feedback significantly influences consumer purchasing decisions. , April 15, 2009. [61] A. Yali, M. Bayram, eWom: The effects of online consumer reviews on purchasing decision of electronic goods, in: Proceedings of the International Marketing Trends Conference, Venice, 2010.

836

C. Kaiser et al. / Knowledge-Based Systems 24 (2011) 824–836

[62] B. Yang, L.X. Li, H. Ji, J. Xu, An early warning system for loan risk assessment using artificial neural networks, Knowledge-Based Systems 14 (5–6) (2001) 303–306. [63] L. Xu, W. He, Application of fuzzy neural network to fire alarm system of highrise building, Journal of Communication and Computer 2 (9) (2005) 18–21.

[64] L. Zadeh, Fuzzy sets, Information and Control 8 (3) (1965) 338–353. [65] W. Zhang, T. Yoshida, X. Tang, Text classification based on multi-word with support vector machine, Knowledge-Based Systems 21 (8) (2008) 879–886. [66] J. Zeng, S. Zhang, C. Wu, A framework for WWW user activity analysis based on user interest, Knowledge-Based Systems 21 (8) (2008) 905–910.