- Email: [email protected]
Feature fatigue analysis in product development using Bayesian networks
Expert Systems with Applications 38 (2011) 10631–10637
Contents lists available at ScienceDirect
Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa
Feature fatigue analysis in product development using Bayesian networks Ming Li, Liya Wang ⇑ Department of Industrial Engineering and Logistics Management, Shanghai Jiao Tong University, No. 800 Dongchuan Road, 200240 Shanghai, China
a r t i c l e
i n f o
Keywords: Feature fatigue analysis Bayesian networks Customer preferences Product development
a b s t r a c t The construct of ‘‘feature fatigue’’ represents the phenomenon of customer’s inconsistent satisfaction: customers prefer to choose products with more features and capacities initially, but once actually worked with a product they will find the complex ones are too hard to use. Clearly, customer’s dissatisfaction after use will have a negative effect on company’s long-term revenue, and the inconsistence is a big challenge for firm’s product development. Researchers have proposed some methods to ‘‘defeat’’ feature fatigue, however, most recent research just analyzes features one by one and ignore the relationships among them. Another problem is that the uncertain nature of customer preferences has not been paid enough attention. To solve these problems, a probability based methodology for feature fatigue analysis is proposed, in which Bayesian network techniques are used to represent the uncertain customer preferences for capacity and usability. And in this method, sensitivity analysis is implemented to identify the key features that affect feature fatigue most, and the relationships among features are analyzed using Bayesian network inference. An example is given to illustrate the usage of the proposed method in product development process. Ó 2011 Elsevier Ltd. All rights reserved.
1. Introduction Intuition and past research suggest that customers usually buy products based on the number of provided features, and they are often seduced by extra features in the moment of purchasing (Kruger, Galak, & Burrus, 2007; Nowlis & Simonson, 1996; Venkatesh & Mahajan, 1993). In today’s competitive environment, satisfying customer needs has become a great concern, and companies try their best to develop products with more features and capacities (Chen & Wang, 2007; Jiao, Zhang, & Helander, 2006). However, many studies show that people are poor predictors of their own enjoyment and happiness, especially when they are facing products with too many features (Mandel & Nowlis, 2008). That means satisfaction and dissatisfaction in the moment of purchasing is not necessarily equal to customers’ experiences after using (Löfgren & Witell, 2008). Actually customers often overestimate the utility of extra features prior to purchasing, and after use, they will complain and even return products considering the problem of usability or mismatch with their expectations (Keijzers, den Ouden, & Lu, 2008). So just focusing on how to attract customers by high-feature products will not be helpful for seller’s long-term revenue.
⇑ Corresponding author. Address: Department of Industrial Engineering and Logistics Management, School of Mechanical Engineering, Shanghai Jiao Tong University, No. 800 Dongchuan Road, 200240 Shanghai, China. Tel.: +86 1381 849 1902. E-mail addresses: [email protected] (M. Li), [email protected] (L. Wang). 0957-4174/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2011.02.126
Many cases have been reported to show this problem. For example, a study points that 63% of mobile phone returns in UK has no hardware or software fault but the reported problems relate to usability like issues about the configuration of the handset (Keijzers et al., 2008). Another case is the BMW 7-Series cars, whose dashboard contains over 700 features. This kind of high capacity car is truly attractive in the first moment, but most of the owners are frustrated by the multi-function displays and multi-step options in the complicated system, and their dissatisfaction will affect BMW’s sale in a long term (Rust, Thompson, & Hamilton, 2006). To represent the phenomenon of customer’s inconsistent satisfaction, Thompson, Hamilton, and Rust (2005) used the construct of ‘‘feature fatigue’’ (FF). Based on some case studies, they indicate that capability and usability are two important factors to affect customer’s long-term satisfaction. When buying products such as a cell phone, even though customers know that too many features will lead to usability problem, they still tend to choose highfeature models because capability gets more attention in this moment. However, after working with a product actually, customers will find that usability becomes more important as products with more features are harder to use. So adding numerous features can increase the perceived capability of loaded product, but at the same time it will reduce the perceived usability. To examine how many features are suitable for a product, Thompson et al. (2005) propose an analytical model considering both before and after use, which can help manager to balance sales benefits and customer usability cost of adding features.
10632
M. Li, L. Wang / Expert Systems with Applications 38 (2011) 10631–10637
To date, the problem of FF has been recognized as an important issue in many fields, and some research has been reported to explain or ‘‘defeat’’ FF. But there are still some limitation and shortcoming in previous research. Facing the problem as which feature should be added to a product considering FF, Thompson et al. (2005) just focus on the total number of features but ignore the difference between them. Some researchers have recognized that different feature has different impact on customer’s perceived capability and usability. Typically they try to divide features into different classes like hedonic and utilitarian ones (Gill, 2008; Tsai & Liu, 2007), but the processes of evaluation and classification are usually one by one, ignoring the relationship analysis among features. In fact, as a product is combined with a set of functions and features, the same feature, no matter hedonic or utilitarian, will have different impact on FF in different combination. For example, if adding FM radio to a cell phone which contains MP3 player function, the perceived capacity may not increase significantly as adding the radio to a simple phone which has no entertainment function. As for usability, customers may not feel the product becomes harder to use if they have paid attention to learn how to use the MP3 player. But for simple phone users who have no related experiences, the FM radio will set the product to a higher complexity level. So when making feature adding decisions, the relationships among features contained in the product should be considered from both capacity and usability aspects. Furthermore, the uncertainty nature of customer preferences is a big challenge for the research of FF. When asking questions like ‘‘How do you feel the capacity of product A’’, customers will give their answers depending on their own experience, feeling, context and even their mood at that time (Corney, 2000). As FF analysis is mainly based on customer preferences for capacity and usability (Rust et al., 2006), the uncertainties must be quantified and combined in the analysis process, to make sure that all decisions are taken on a rational basis. It is, of course, an extremely complex task because of the large number of features and their relationships to be considered and the large uncertainty associated with customer preferences for the whole product and each individual feature. To overcome the existing problems, we propose a probability based methodology to deal with the task of FF analysis in product development. In this method Bayesian network is used to represent the structure of a product. And we will describe how to build and use Bayesian network based on customers’ perceived capability and usability data. The proposed methodology will address the issues concerned with FF embedded with uncertainties of customer preferences and will help decision-makers to make intelligent decisions during the process of product development. The remainder of this paper proceeds as follows. In the next section, we elaborate on the recent literature on FF. In Section 3, the proposed method is introduced. And an example is presented to illustrate the usage of this method in Section 4. Finally, in Section 5, conclusions are given.
product in practice, customers realize that the problem of usability is much more important than they ever thought, especially when too many features make the product too complex to use. The difference of customer’s utility function for high-feature product before and after use will lead to dissatisfaction and the phenomenon of FF, which will have a negative effect on company’s longterm revenue. To solve this problem Thompson et al. (2005) propose an analytical model to decide the suitable number of a product’s features considering customers’ reaction over a long period of time. To explain the reason for FF, Hamilton and Thompson (2007) use construal level theory to compare the different effect of direct and indirect product experiences. Their studies show that direct experiences trigger more concrete mental construal and increases preferences of high feasibility/low desirability products, while customers who engage in an indirect experience prefer high desirability/low feasibility products which will lead to the situation of FF. As a shift in construal is the mechanism responsible for the change in preferences, firms are suggested to increase experiential contact with products before purchase by opportunities for product testing. Advertising and online shopping are also helpful to encourage customers thinking concretely about the product they want to purchase. Researchers have recognized the problem of FF in many fields. Tep (2009) use the FF notion in the context of web site’s affective quality and features quality in relation to customer online satisfaction. Tsai and Liu (2007) explore the phenomenon of FF in customer decision making under the scenarios of forfeiture and acquisition, and examine customer’s inequality of weights on utilitarian and hedonic features in the process of buying a car. Keijzers et al. (2008) discuss the usability problem of smart phone which is a typical high-feature product. In order to ‘‘defeat’’ FF, Thompson et al. (2005) focus on how to decide the suitable number of features, and suggest that firms should offer a wider assortment of simpler products instead of all-purpose and feature-rich products. Rahman and Rahman (2009) argue that reducing the number of feature is unhealthy considering the presence of competition, and product variety will impose customers’ search cost. Keijzers et al. (2008) also point that just restrict new product to fewer functions and features will not helpful and a trade-off has to be made between marketing, design and internal quality costs. Another problem is different features will have different effect on FF. Recent research have shown that customers not only place importance on the amount of the total features, but also the amount of different group of features like utilitarian attributes and hedonic ones (Gill, 2008; Tsai & Liu, 2007). In this paper, we propose a probability based methodology for feature fatigue analysis. Not like previous study, we consider features’ relationships in combination rather than one by one, and Bayesian network techniques are used to reflect the uncertainty nature of customer preferences. The proposed method is introduced in the next section.
2. Literature review FF is used to represent the phenomenon that customers are overwhelmed by too many product features. Traditionally, market and economic research indicates that more features will make product more appealing to customers, and company’s profit will increase at the same time (Ellison, 2005; Green & Krieger, 1991; Hauser & Rao, 2003). It is true that when a product concludes more features, its perceived capacity will increase, but the perceived usability will decrease considering the problem of complexity. Thompson et al. (2005) show that at the point of purchasing, customers prefer to choose high-feature products, that means they give more weight to the factor of capacity. But when using the
3. Bayesian network based FF analysis method A Bayesian network is a kind of powerful knowledge representation and reasoning tools under conditions of uncertainty (Heckerman, 1996). Formally, a Bayesian network consists of three parts B ¼ hV; A; Pi. The first part V is a set of discrete and stochastic variables. In a network, the variables are represented by nodes. The second part A is a set of arcs which connect the nods and indicate direct dependencies between the variables. V and A constitute a direct acyclic graph G ¼ hV; Ai. The last part P represents a set of conditional probability distribution. Suppose V represents
M. Li, L. Wang / Expert Systems with Applications 38 (2011) 10631–10637
the variables X1, . . . , Xn, the probability distribution P can be calculated as
PðX 1 ; . . . ; X n Þ ¼
n Y
PðX i jpðX i ÞÞ
ð1Þ
i¼1
where p(Xi) denotes the set of direct parents of Xi in G. Comparing with other analysis approaches, Bayesian network techniques have a number of advantages that makes them suitable for the FF problem. Firstly, determinate methods are hard to catch the uncertainty nature of customer preferences (Wang & Tseng, 2007), while Bayesian network can represent the probability of an event x as a person’s degree of belief in that event, and this feature makes it possible to efficiently handle the uncertainty problem of customer preferences for capability and usability. Secondly, in a Bayesian network the probability distribution for each variable depends only on the node’s parents, so fewer parameters need to be estimated. This is an important advantage when too many variables are contained in a network, and for our problem, that means usually a large number of features are contained in a high-feature product. Thirdly, many other intelligent systems (such as fuzzy logic and feed-forward neural network) are strictly one-way. Given a set of inputs the one-way methods can predict related outputs, but they cannot answer questions like ‘‘What a product will be like if its capacity level is high?’’ But as a bi-direction inference method Bayesian network can handle this kind of inference problem inversely (Lu, Bai, & Zhang, 2009). The fourth advantage is that Bayesian network has a causal semantics that makes the encoding of causal prior knowledge particularly straightforward. This feature helps users to construct a Bayesian network using their domain knowledge especially when data is scare or expensive. The features talking above show that Bayesian network is a suitable method to verify those initially identified uncertain relationships between product features. So in this paper a Bayesian network based FF analysis model is proposed and the detailed descriptions of how to construct and use this model are shown as follows. 3.1. Capability and complexity measurement for a Bayesian network As shown above, FF represents this kind of phenomenon: consumers prefer to choose products with more features in the moment of purchasing and then confused and frustrated after they actually using these high-feature products. Thompson et al. (2005) indicate that if firms want to defeat FF they should balance the initial sales and long-term customer satisfaction. In one word, capacity and usability are the two factors that reflect the effect of FF on company’s long-term revenue. In this paper, we use complexity to represent the lack of usability, and customers are asked to evaluate the capacity and complexity level of the whole product and each of its features. We assume that there are three levels: L (low), M (moderate) and H (high). For example, when choosing a cell phone a customer may find MP3 function attract him or her most, so the capacity level of MP3 feature can be ‘‘H’’. When actually using this cell phone, the consumer may find the MP3 function is very easy to use, so the complexity level can be ‘‘L’’. As for some basic functions like message sending, consumers usually give a low capacity level, but if they find sending message is hard to use they will be dissatisfied and give a high complexity level. These survey data are collected for Bayesian network building in next stage.
10633
1997). Firstly, the graphical Bayesian network structure should be defined. Then the related parameters in the form of conditional probabilities need to be calculated. The structure of a Bayesian network contains a set of nodes represent random variables and arcs depict direct probabilistic dependences among them. In some cases, the structure can be generated through knowledge elicitation from domain experts. But usually the manual expert-based method is time-consuming because the entire structure should be defined by hand, and in many cases the suitable domain knowledge is not available (Corney, 2000). Another method is to learn Bayesian network structure from dada. But as a function of the number of nodes, the number of possible structures grows exponentially. So the exhaustive enumeration of all network structures is not feasible (Cooper & Herskovits, 1992). To solve this problem a number of search algorithms have been proposed. In this paper, the Greedy Thick Thinning algorithm (GTT) algorithm will be used to construct the Bayesian network for FF analysis. Generally the GTT algorithm works in two stages. Firstly, it starts with an empty graph and adds (or reverses) the arc (without creating a cycle) that increases the score as much as possible. This process is repeated until no adding or reversing can increase the score anymore. After that it repeatedly removes arcs until no arc deletion will result in a positive increase in the score (Dash & Druzdzel, 2003). Once the structure has been constructed, parameters (probabilities for each node) should be decided next. The easiest way of calculating the probabilities is using data frequency. However, as the size of data in the FF analysis study is usually not very large, using a frequency method will be not very suitable, as many feature combinations will not have been observed. To solve this problem, the Expectation Maximization (EM) algorithm will be used. The EM method can fill in the missing values with the most likely values as if the dataset is complete, and then it will iterate until convergence (Dempster, Laird, & Rubin, 1977). 3.3. FF analysis Now two Bayesian networks, one for capacity analysis and the other for complexity (lack of usability), have been constructed defining both the structure and conditional probabilities. We can use the Bayesian networks to conduct inference to the relationships among features and the whole product. To make inference we should firstly fix the states of known variables, and then propagate the beliefs around the network until all beliefs are consistent. After that we can get the desired probability distribution directly from the network. Belief updating is computationally complex, and in the worst case belief updating algorithms are NP-hard (Cooper, 1990). In this paper we use clustering algorithm proposed by Lauritzen and Spiegelhalter (1988) to deal with the updating task. This algorithm is fast and efficient to make belief updating in Bayesian network with hundreds of variables tractable. To find the key features which affect FF most, some concepts and techniques from information theory are used to deal with the analysis work. The concept of entropy in information theory was first introduced by Shannon (1948). Entropy is a measure of the uncertainty associated with a random variable. Let X be a discrete random variable, given value x, the probability distribution function is p(x) = Pr(X = x). The entropy H(X) of X can be defined as
HðXÞ ¼
X
pðxÞ log pðxÞ
ð2Þ
x2X
3.2. Bayesian network construction After getting the dataset of FF discussed in Section 3.1, the related Bayesian network will be constructed in this stage. Typically, there are two steps to build a Bayesian network (Heckerman,
Note that the entropy of X does not depend on the actual values of X, it only depends on p(x). Suppose X and Y are two random variables. Then for any fixed value y of Y, we can get a conditional probability distribution p(X|y) as follows:
10634
HðXjyÞ ¼
M. Li, L. Wang / Expert Systems with Applications 38 (2011) 10631–10637
X
pðxjyÞ logðpðxjyÞÞ
ð3Þ
x2X
Let the conditional entropy H(X|Y) be defined to be the expectation of H(X|y) when average over all possible values y. Then it is
HðXjYÞ ¼
XX
pðyÞpðxjyÞ logðpðxjyÞÞ
when using complex products, this criteria actually represents the comprehensive effect of FF in the view of companies. And the result of Fx minus Fa reflects the FF degree of the related product, and for FFD the lower the better. A more detailed analysis will be shown in the following example.
ð4Þ
y2Y x2X
4. Example case
Now to measure the mutual information of X and Y, I(X|Y) can be defined as
IðXjYÞ ¼ HðXÞ HðXjYÞ
ð5Þ
This measure represents the difference of the a priori and a posteriori entropies of X, i.e., the reduction in uncertainty about X by knowing Y. It is noted that I(X|Y) is the information in X about Y, and a high degree of dissimilarity indicates more information X carries about Y. As its name suggests, mutual information is symmetric, so I(X|Y) = I(Y|X). Entropy reduction is suitable for sensitivity analysis in Bayesian network (Lee, Park, & Shin, 2009), and in this paper we use it to identify the key features that affect FF most. After the key features have been selected, experts in the company can determine these features’ capacity and complexity level from both marketing and engineering concerns. Usually some candidate products will be given and their FF degree can be measured as
FFD ¼ Fx Fa
ð6Þ
where Fx is the complexity score, Fa is the capacity score. As Thompson et al. (2005) shown that, to defeat FF firms should balance the initial sales and long-term customer satisfaction. So in Eq. (6), FFD does not just reflect the fatigue feeling of consumers Table 1 Probabilities of the features in the capacity Bayesian network. Feature
HA
MA
LA
Digital camera Games GPS Wireless internet FM radio Email MP3 player Bluetooth
81.9 69.0 62.0 60.1 58.1 44.2 30.0 16.5
11.9 21.0 26.0 16.2 26.6 43.9 58.7 73.7
6.17 10.0 12.0 23.6 15.3 12.0 11.2 9.78
4.1. Case description To illustrate the approaches proposed in this paper, an example of FF analysis for smart phone development is presented in this section. Smart phones are typical high-feature products as manufacturers are lured into integration of a growing number of technologies and features to provide attractive and competitive models (Zheng & Ni, 2006). Customers initially choose these high-feature smart phones but after use they will find that the usability of such complex products becomes an increasing problem (Keijzers et al., 2008). In this paper eight major features of smart phone are selected to analyze their different effect on FF in the following paragraphs. 4.2. Capability and complexity measurement The selected features are listed in Table 1. And we also get 100 customers’ evaluations. As discussed in Section 3.1, the FF problem is directly related to two factors: capacity and complexity. To simulate the real market situation, the 100 ‘‘customers’’ are asked to respectively evaluate the capacity and complexity level of the eight features and the whole product. The capacity evaluation represents customer preferences at the point of purchasing, and after use the direct experience will affect the complexity level. A 3-level scale is used to present the evaluation: H (high), M (moderate) and L (low). For capacity the three levels are HA, MA and LA, and for complexity they are HX, MX and LX. 4.3. Bayesian network construction In this section two Bayesian networks are constructed based on the evaluation data. For each network GTT algorithm is firstly used to construct the structure and then EM algorithm is used for
Fig. 1. Bayesian network for capacity analysis.
10635
M. Li, L. Wang / Expert Systems with Applications 38 (2011) 10631–10637
Fig. 2. Bayesian network for complexity analysis.
parameter learning. The models described in this paper are created using the GeNIe modeling environment, which is a type of Bayesian network and decision support software and developed by the Decision Systems Laboratory of the University of Pittsburgh (http://dsl.sis.pitt.edu). The two networks for capacity and complexity evaluation are respectively shown in Figs. 1 and 2, in which FFA represent the capacity level of the whole product, and FFX means the product’s complexity level. 4.4. FF analysis for smart phone development Given the complete Bayesian network with both structure and conditional probabilities, we can make inferences and sensitivity analysis to guide the task of defeating FF. Firstly, Table 1 shows the probabilities of the features in Fig. 1 ordering by HA, which is the highest level of capacity. And Table 2 represents the probabilities in Fig. 2 ordering by the highest complexity level HX. Note that these probabilities result from the Bayesian network calculations rather than frequency data. From Eq. (6), we know a feature with low capacity level and high complexity level will have a high FF degree. Just from the result of Tables 1 and 2 the features like Bluetooth (HA = 16.5% and HX = 32.2%) should be paid attention to improve their capacity and/or reduce their complexity. But this simple conclusion ignores the relationships among features and their different effect on the whole product’s FF degree. More analysis will be made in the following paragraphs. As discussed in Section 3.3, entropy reflects the relative influence between uncertain variables. Tables 3 and 4 show the result of entropy reduction calculation from sensitivity analysis of
features related to FFA and FFX respectively. If set 0.05 as the threshold value for decision, the features influence FFA most are Wireless internet, games, MP3 player, Email and digital camera. And the features of MP3 player, digital camera and wireless internet have the greatest influence on FFX. The sensitivity analysis can help to identify the most influential decisions to defeat FF. For example, the Wireless internet feature has a much larger entropy reduction value than the feature of Bluetooth for both FFA and FFX. That means Wireless internet can provide more information and eliminate more uncertainty of FFA and FFX, while trying to confirm the Bluetooth feature will have less effect on the probability distribution of the whole product’s capacity and complexity.
Table 3 Entropy reduction values in the capacity Bayesian network. Feature
I(X; Y)
Wireless internet Games MP3 player Email Digital camera FM radio GPS Bluetooth
0.20756 0.15981 0.08770 0.07562 0.06231 0.02502 0.01056 0.00835
Table 4 Entropy reduction values in the complexity Bayesian network.
Table 2 Probabilities of the features in the complexity Bayesian network. Feature
HX
MX
LX
Feature
I(X; Y)
GPS Email Games Wireless internet MP3 player Bluetooth Digital camera FM radio
49.0 46.0 39.1 39.0 38.0 32.2 26.7 18.0
22.0 19.0 24.9 16.0 40.4 28.5 46.3 66.0
29.0 35.0 36.0 45.0 21.6 39.3 27.0 16.0
MP3 player Digital camera Wireless internet FM radio Email Bluetooth GPS Games
0.21741 0.16908 0.07478 0.01829 0.01698 0.01573 0.00359 0.00247
10636
M. Li, L. Wang / Expert Systems with Applications 38 (2011) 10631–10637
Table 5 FF evaluation results for candidate products. Figure
P1
P2
Capacity
Complexity
Capacity
Complexity
Digital camera MP3 player
HA = 100% MA = 100%
HX = 100% MX = 100%
MA = 100% HA = 100%
MX = 100% HX = 100%
Other features
Initial probability distribution
FFA FFX
HA = 41.5%
HA = 33.3% HX = 24.9%
HX = 23.3%
Based on the entropy reduction results, experts in the company can choose key features affect product’s FF most and decide their capacity and complexity level from both marketing and engineering concerns. Usually some candidates will be given in this step and the next task is how to evaluate them. In a Bayesian network, if the values of some variables are given, the probabilities of the remaining variables can be calculated. In this case, if one or some key features’ capacity and complexity level have been decided, the probability of FFA in Fig. 1 and FFX in Fig. 2 will be changed and can be calculated using the related network. An example of evaluation is shown in Table 5. P1 and P2 are two candidate products. For P1 the digital camera (DC) feature has a high capacity level HA and high complexity level HX. That means the DC feature is very attractive but hard to use. And the MP3 player of P1 is moderate for both capacity and complexity. Other features’ capability and complexity level is not decided and given the initial probability distribution. Using Figs. 1 and 2 the probability of high capacity level HA for P1 is 41.5%, and HX is 24.9%. For P2, its HA is 33.3% and HX is 23.3%. To evaluate these two products, Eq. (6) should be used. To be simple, we use the probability of HA to represent the capacity score Fa, and HX’s probability as the complexity score Fx. So the FF degree of P1 is 0.166 and P2 is 0.1. As P1 has a lower FF degree, it is better than P2 considering both capability and complexity. From Table 5 we can see even the probability of P1’s HX (24.9%) is a little larger than P2’s (23.3%), but P1 is much more attractive than P2 (41.5% versus 23.3%). Besides sensitivity analysis and evaluation presented above, Bayesian network can also be used to analyze the relationships among features, and this is useful in the process of deciding key features’ capacity and complexity level. Figs. 3 and 4 show the effect of setting the value of MP3. In Fig. 3 the first bar of each pair represents the prior probability of the variable having a ‘‘HA’’ value, and the second bar is the corresponding posterior probability after the value of some variables like MP3 has been decided. And Fig. 4 shows the probability change of HX. We can find that when the value of MP3 is ‘‘HA’’, the probability of FM’s HA value has increased from 0.581 to 0.721. This increase indicates that the capacity of MP3 and FM are correlated to some extent, that is, a HA MP3
Fig. 4. Prior and posterior probabilities of HX when set MP3 to HX.
tends to ‘‘cause’’ the increase of FM’s HA level probability. We can also find that the FFA increase from 0.511 to 0.753. That means a high capacity MP3 will also bring a significant capacity enhancement of the product. The same process can also be used for the analysis of complexity relationships among different features in Fig. 4. 5. Conclusion In this paper, a probability based methodology was proposed to analyze the effect of FF on company’s long-term revenue. Bayesian network techniques were used to represent the uncertain customer preferences for capacity and usability. In order to construct a Bayesian network, customers were asked to evaluate the capacity and complexity (lack of usability) level of the whole product and each individual feature. Based on these evaluation data Bayesian networks were built and used for the FF analysis task. Sensitivity analysis was implemented based on entropy reduction to identify the key features that affect product’s FF most. Experts in the company could determine the capacity and complexity level of these key features from both marketing and engineering concerns. After that some candidate products were constructed, and their FF degree could be got through calculation based on the Bayesian networks. And using inference techniques the relationships among features were analyzed by comparing the prior and posterior probabilities. The method proposed in this paper can guide decision-makers in marketing and engineering to make the most influential decisions in the process of product development. As the application of probabilistic graphical models, the uncertainty problem of customer preferences for capability and complexity could be handled efficiently. If enough data are hard to get in some situations, the Bayesian network can be constructed through experts’ domain knowledge and the proposed method can still work for FF analysis. As a closing remark, we are aware of the limitations within this study. One is that to build a Bayesian network, the features contained in a product should be given first. If decision-makers want to delete/add one or some features, they should rebuild Bayesian network from the initial data gathering step. Another limitation is that the data used in this study just focuses on ‘‘present’’ customer preferences. In fact, customer preferences are dynamic and may vary drastically from time to time (Chen & Wang, 2008). So our future work is to be aimed at dealing with the problem of changeable features and dynamic customer preferences in the process of defeating FF. Acknowledgements
Fig. 3. Prior and posterior probabilities of HA when set MP3 to HA.
This research is supported by the National Natural Science Foundation of China (Grant No. 70932004/G0209), and the
M. Li, L. Wang / Expert Systems with Applications 38 (2011) 10631–10637
National Natural Science Foundation of China (Grant No. 71072061/G020801). References Chen, Z. X., & Wang, L. Y. (2007). Integrating knowledge acquisition and deduction in product configuration by neural networks. In The third IEEE international conference on wireless communications, networking and mobile computing. Chen, C., & Wang, L. (2008). Integrating rough set clustering and grey model to analyse dynamic customer requirements. Proceedings of the Institution of Mechanical Engineers, Part B: Journal of Engineering Manufacture, 222(2), 319–332. Cooper, G. F. (1990). Research note the computational complexity of probabilistic inference using Bayesian belief networks. Artificial Intelligence, 42, 393–405. Cooper, G. F., & Herskovits, E. (1992). A Bayesian method for the induction of probabilistic networks from data. Machine Learning, 9(4), 309–347. Corney, D. (2000). Designing food with Bayesian belief networks. Adaptive computing in design and manufacture. Univ. of Plymouth. pp. 83–94. Dash, D., & Druzdzel, M. J. (2003). Robust independence testing for constraint-based learning of causal structure. Proceedings of the nineteenth conference on uncertainty in artificial intelligence, 167–174. Dempster, A. P., Laird, N. M., & Rubin, D. B. (1977). Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical Society. Series B (Methodological), 39(1), 1–38. Ellison, G. D. (2005). A model of add-on pricing. The Quarterly Journal of Economics, 120(2), 585–637. Gill, T. (2008). Convergent products: What functionalities add more value to the base? Journal of Marketing, 72(2), 46–62. Green, P. E., & Krieger, A. M. (1991). Segmenting markets with conjoint analysis. The Journal of Marketing, 20–31. Hamilton, R. W., & Thompson, D. V. (2007). Is there a substitute for direct experience? Comparing consumers’ preferences after direct and indirect product experiences. Journal of Consumer Research, 34(4), 546–555. Hauser, J. R., & Rao, V. R. (2003). Conjoint analysis, related modeling, and applications. Marketing Research and Modeling: Progress and Prospects: A Tribute to Paul Green, 141–168. Heckerman, D. (1996). A tutorial on learning Bayesian networks. Technical Report MSRTR-95-06, Microsoft Research. Heckerman, D. (1997). Bayesian networks for data mining. Data Mining and Knowledge Discovery, 1(1), 79–119. Jiao, J., Zhang, Y., & Helander, M. (2006). A Kansei mining system for affective design. Expert Systems with Applications, 30(4), 658–673.
10637
Keijzers, J., den Ouden, E., & Lu, Y. (2008). Usability benchmark study of commercially available smart phones: Cell phone type platform, PDA type platform and PC type platform. In Proceedings of the 10th international conference on Human computer interaction with mobile devices and services (pp. 265–272). Kruger, J., Galak, J., & Burrus, J. (2007). When consumers’ self-image motives fail. Journal of Consumer Psychology, 17(4), 250–253. Lauritzen, S. L., & Spiegelhalter, D. J. (1988). Local computations with probabilities on graphical structures and their application to expert systems. Journal of the Royal Statistical Society. Series B (Methodological), 50(2), 157–224. Lee, E., Park, Y., & Shin, J. G. (2009). Large engineering project risk management using a Bayesian belief network. Expert Systems with Applications, 36(3P2), 5880–5887. Löfgren, M., & Witell, L. (2008). Two decades of using Kano’s theory of attractive quality: A literature review. The Quality Management Journal, 15(1), 59. Lu, J., Bai, C., & Zhang, G. (2009). Cost-benefit factor analysis in e-services using Bayesian networks. Expert Systems with Applications, 36(3P1), 4617–4625. Mandel, N., & Nowlis, S. M. (2008). The effect of making a prediction about the outcome of a consumption experience on the enjoyment of that experience. Journal of Consumer Research, 35(1), 9–20. Nowlis, S. M., & Simonson, I. (1996). The effect of new product features on brand choice. Journal of Marketing Research, 33(February), 36–46. Rahman, M., & Rahman, M. M. (2009). To defeat feature fatigue the right way, understand it first. Strategic Direction, 25(6), 26–28. Rust, R. T., Thompson, D. V., & Hamilton, R. W. (2006). Defeating feature fatigue. Harvard Business Review, 84(2), 98. Shannon, C. E. (1948). A mathematical theory of communication Reprinted with corrections from. The Bell System Technical Journal, 27(379–423), 623–656. Tep, S.-B. S. P. (2009). Web site affective quality and features quantity in relation to customer online satisfaction: A research framework. In The 3rd student conference on business research. Thompson, D. V., Hamilton, R. W., & Rust, R. T. (2005). Feature fatigue: When product capabilities become too much of a good thing. Journal of Marketing Research, 42(4), 431–442. Tsai, H. M., & Liu, C.H. (2007). The influence of feature fatigue on the preferences of hedonic and utilitarian features. In International conference on business and information. Venkatesh, R., & Mahajan, V. (1993). A probabilistic approach to pricing a bundle of products or services. Journal of Marketing Research, 30(November), 494–508. Wang, Y. & Tseng, M. M. (2007). An approach to improve the efficiency of configurators. In 2007 IEEE international conference on industrial engineering and engineering management (pp. 1332–1336). Zheng, P., & Ni, L. M. (2006). Smart phone and next generation mobile computing. San Francisco: Morgan Kaufmann.
Contents lists available at ScienceDirect
Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa
Feature fatigue analysis in product development using Bayesian networks Ming Li, Liya Wang ⇑ Department of Industrial Engineering and Logistics Management, Shanghai Jiao Tong University, No. 800 Dongchuan Road, 200240 Shanghai, China
a r t i c l e
i n f o
Keywords: Feature fatigue analysis Bayesian networks Customer preferences Product development
a b s t r a c t The construct of ‘‘feature fatigue’’ represents the phenomenon of customer’s inconsistent satisfaction: customers prefer to choose products with more features and capacities initially, but once actually worked with a product they will find the complex ones are too hard to use. Clearly, customer’s dissatisfaction after use will have a negative effect on company’s long-term revenue, and the inconsistence is a big challenge for firm’s product development. Researchers have proposed some methods to ‘‘defeat’’ feature fatigue, however, most recent research just analyzes features one by one and ignore the relationships among them. Another problem is that the uncertain nature of customer preferences has not been paid enough attention. To solve these problems, a probability based methodology for feature fatigue analysis is proposed, in which Bayesian network techniques are used to represent the uncertain customer preferences for capacity and usability. And in this method, sensitivity analysis is implemented to identify the key features that affect feature fatigue most, and the relationships among features are analyzed using Bayesian network inference. An example is given to illustrate the usage of the proposed method in product development process. Ó 2011 Elsevier Ltd. All rights reserved.
1. Introduction Intuition and past research suggest that customers usually buy products based on the number of provided features, and they are often seduced by extra features in the moment of purchasing (Kruger, Galak, & Burrus, 2007; Nowlis & Simonson, 1996; Venkatesh & Mahajan, 1993). In today’s competitive environment, satisfying customer needs has become a great concern, and companies try their best to develop products with more features and capacities (Chen & Wang, 2007; Jiao, Zhang, & Helander, 2006). However, many studies show that people are poor predictors of their own enjoyment and happiness, especially when they are facing products with too many features (Mandel & Nowlis, 2008). That means satisfaction and dissatisfaction in the moment of purchasing is not necessarily equal to customers’ experiences after using (Löfgren & Witell, 2008). Actually customers often overestimate the utility of extra features prior to purchasing, and after use, they will complain and even return products considering the problem of usability or mismatch with their expectations (Keijzers, den Ouden, & Lu, 2008). So just focusing on how to attract customers by high-feature products will not be helpful for seller’s long-term revenue.
⇑ Corresponding author. Address: Department of Industrial Engineering and Logistics Management, School of Mechanical Engineering, Shanghai Jiao Tong University, No. 800 Dongchuan Road, 200240 Shanghai, China. Tel.: +86 1381 849 1902. E-mail addresses: [email protected] (M. Li), [email protected] (L. Wang). 0957-4174/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2011.02.126
Many cases have been reported to show this problem. For example, a study points that 63% of mobile phone returns in UK has no hardware or software fault but the reported problems relate to usability like issues about the configuration of the handset (Keijzers et al., 2008). Another case is the BMW 7-Series cars, whose dashboard contains over 700 features. This kind of high capacity car is truly attractive in the first moment, but most of the owners are frustrated by the multi-function displays and multi-step options in the complicated system, and their dissatisfaction will affect BMW’s sale in a long term (Rust, Thompson, & Hamilton, 2006). To represent the phenomenon of customer’s inconsistent satisfaction, Thompson, Hamilton, and Rust (2005) used the construct of ‘‘feature fatigue’’ (FF). Based on some case studies, they indicate that capability and usability are two important factors to affect customer’s long-term satisfaction. When buying products such as a cell phone, even though customers know that too many features will lead to usability problem, they still tend to choose highfeature models because capability gets more attention in this moment. However, after working with a product actually, customers will find that usability becomes more important as products with more features are harder to use. So adding numerous features can increase the perceived capability of loaded product, but at the same time it will reduce the perceived usability. To examine how many features are suitable for a product, Thompson et al. (2005) propose an analytical model considering both before and after use, which can help manager to balance sales benefits and customer usability cost of adding features.
10632
M. Li, L. Wang / Expert Systems with Applications 38 (2011) 10631–10637
To date, the problem of FF has been recognized as an important issue in many fields, and some research has been reported to explain or ‘‘defeat’’ FF. But there are still some limitation and shortcoming in previous research. Facing the problem as which feature should be added to a product considering FF, Thompson et al. (2005) just focus on the total number of features but ignore the difference between them. Some researchers have recognized that different feature has different impact on customer’s perceived capability and usability. Typically they try to divide features into different classes like hedonic and utilitarian ones (Gill, 2008; Tsai & Liu, 2007), but the processes of evaluation and classification are usually one by one, ignoring the relationship analysis among features. In fact, as a product is combined with a set of functions and features, the same feature, no matter hedonic or utilitarian, will have different impact on FF in different combination. For example, if adding FM radio to a cell phone which contains MP3 player function, the perceived capacity may not increase significantly as adding the radio to a simple phone which has no entertainment function. As for usability, customers may not feel the product becomes harder to use if they have paid attention to learn how to use the MP3 player. But for simple phone users who have no related experiences, the FM radio will set the product to a higher complexity level. So when making feature adding decisions, the relationships among features contained in the product should be considered from both capacity and usability aspects. Furthermore, the uncertainty nature of customer preferences is a big challenge for the research of FF. When asking questions like ‘‘How do you feel the capacity of product A’’, customers will give their answers depending on their own experience, feeling, context and even their mood at that time (Corney, 2000). As FF analysis is mainly based on customer preferences for capacity and usability (Rust et al., 2006), the uncertainties must be quantified and combined in the analysis process, to make sure that all decisions are taken on a rational basis. It is, of course, an extremely complex task because of the large number of features and their relationships to be considered and the large uncertainty associated with customer preferences for the whole product and each individual feature. To overcome the existing problems, we propose a probability based methodology to deal with the task of FF analysis in product development. In this method Bayesian network is used to represent the structure of a product. And we will describe how to build and use Bayesian network based on customers’ perceived capability and usability data. The proposed methodology will address the issues concerned with FF embedded with uncertainties of customer preferences and will help decision-makers to make intelligent decisions during the process of product development. The remainder of this paper proceeds as follows. In the next section, we elaborate on the recent literature on FF. In Section 3, the proposed method is introduced. And an example is presented to illustrate the usage of this method in Section 4. Finally, in Section 5, conclusions are given.
product in practice, customers realize that the problem of usability is much more important than they ever thought, especially when too many features make the product too complex to use. The difference of customer’s utility function for high-feature product before and after use will lead to dissatisfaction and the phenomenon of FF, which will have a negative effect on company’s longterm revenue. To solve this problem Thompson et al. (2005) propose an analytical model to decide the suitable number of a product’s features considering customers’ reaction over a long period of time. To explain the reason for FF, Hamilton and Thompson (2007) use construal level theory to compare the different effect of direct and indirect product experiences. Their studies show that direct experiences trigger more concrete mental construal and increases preferences of high feasibility/low desirability products, while customers who engage in an indirect experience prefer high desirability/low feasibility products which will lead to the situation of FF. As a shift in construal is the mechanism responsible for the change in preferences, firms are suggested to increase experiential contact with products before purchase by opportunities for product testing. Advertising and online shopping are also helpful to encourage customers thinking concretely about the product they want to purchase. Researchers have recognized the problem of FF in many fields. Tep (2009) use the FF notion in the context of web site’s affective quality and features quality in relation to customer online satisfaction. Tsai and Liu (2007) explore the phenomenon of FF in customer decision making under the scenarios of forfeiture and acquisition, and examine customer’s inequality of weights on utilitarian and hedonic features in the process of buying a car. Keijzers et al. (2008) discuss the usability problem of smart phone which is a typical high-feature product. In order to ‘‘defeat’’ FF, Thompson et al. (2005) focus on how to decide the suitable number of features, and suggest that firms should offer a wider assortment of simpler products instead of all-purpose and feature-rich products. Rahman and Rahman (2009) argue that reducing the number of feature is unhealthy considering the presence of competition, and product variety will impose customers’ search cost. Keijzers et al. (2008) also point that just restrict new product to fewer functions and features will not helpful and a trade-off has to be made between marketing, design and internal quality costs. Another problem is different features will have different effect on FF. Recent research have shown that customers not only place importance on the amount of the total features, but also the amount of different group of features like utilitarian attributes and hedonic ones (Gill, 2008; Tsai & Liu, 2007). In this paper, we propose a probability based methodology for feature fatigue analysis. Not like previous study, we consider features’ relationships in combination rather than one by one, and Bayesian network techniques are used to reflect the uncertainty nature of customer preferences. The proposed method is introduced in the next section.
2. Literature review FF is used to represent the phenomenon that customers are overwhelmed by too many product features. Traditionally, market and economic research indicates that more features will make product more appealing to customers, and company’s profit will increase at the same time (Ellison, 2005; Green & Krieger, 1991; Hauser & Rao, 2003). It is true that when a product concludes more features, its perceived capacity will increase, but the perceived usability will decrease considering the problem of complexity. Thompson et al. (2005) show that at the point of purchasing, customers prefer to choose high-feature products, that means they give more weight to the factor of capacity. But when using the
3. Bayesian network based FF analysis method A Bayesian network is a kind of powerful knowledge representation and reasoning tools under conditions of uncertainty (Heckerman, 1996). Formally, a Bayesian network consists of three parts B ¼ hV; A; Pi. The first part V is a set of discrete and stochastic variables. In a network, the variables are represented by nodes. The second part A is a set of arcs which connect the nods and indicate direct dependencies between the variables. V and A constitute a direct acyclic graph G ¼ hV; Ai. The last part P represents a set of conditional probability distribution. Suppose V represents
M. Li, L. Wang / Expert Systems with Applications 38 (2011) 10631–10637
the variables X1, . . . , Xn, the probability distribution P can be calculated as
PðX 1 ; . . . ; X n Þ ¼
n Y
PðX i jpðX i ÞÞ
ð1Þ
i¼1
where p(Xi) denotes the set of direct parents of Xi in G. Comparing with other analysis approaches, Bayesian network techniques have a number of advantages that makes them suitable for the FF problem. Firstly, determinate methods are hard to catch the uncertainty nature of customer preferences (Wang & Tseng, 2007), while Bayesian network can represent the probability of an event x as a person’s degree of belief in that event, and this feature makes it possible to efficiently handle the uncertainty problem of customer preferences for capability and usability. Secondly, in a Bayesian network the probability distribution for each variable depends only on the node’s parents, so fewer parameters need to be estimated. This is an important advantage when too many variables are contained in a network, and for our problem, that means usually a large number of features are contained in a high-feature product. Thirdly, many other intelligent systems (such as fuzzy logic and feed-forward neural network) are strictly one-way. Given a set of inputs the one-way methods can predict related outputs, but they cannot answer questions like ‘‘What a product will be like if its capacity level is high?’’ But as a bi-direction inference method Bayesian network can handle this kind of inference problem inversely (Lu, Bai, & Zhang, 2009). The fourth advantage is that Bayesian network has a causal semantics that makes the encoding of causal prior knowledge particularly straightforward. This feature helps users to construct a Bayesian network using their domain knowledge especially when data is scare or expensive. The features talking above show that Bayesian network is a suitable method to verify those initially identified uncertain relationships between product features. So in this paper a Bayesian network based FF analysis model is proposed and the detailed descriptions of how to construct and use this model are shown as follows. 3.1. Capability and complexity measurement for a Bayesian network As shown above, FF represents this kind of phenomenon: consumers prefer to choose products with more features in the moment of purchasing and then confused and frustrated after they actually using these high-feature products. Thompson et al. (2005) indicate that if firms want to defeat FF they should balance the initial sales and long-term customer satisfaction. In one word, capacity and usability are the two factors that reflect the effect of FF on company’s long-term revenue. In this paper, we use complexity to represent the lack of usability, and customers are asked to evaluate the capacity and complexity level of the whole product and each of its features. We assume that there are three levels: L (low), M (moderate) and H (high). For example, when choosing a cell phone a customer may find MP3 function attract him or her most, so the capacity level of MP3 feature can be ‘‘H’’. When actually using this cell phone, the consumer may find the MP3 function is very easy to use, so the complexity level can be ‘‘L’’. As for some basic functions like message sending, consumers usually give a low capacity level, but if they find sending message is hard to use they will be dissatisfied and give a high complexity level. These survey data are collected for Bayesian network building in next stage.
10633
1997). Firstly, the graphical Bayesian network structure should be defined. Then the related parameters in the form of conditional probabilities need to be calculated. The structure of a Bayesian network contains a set of nodes represent random variables and arcs depict direct probabilistic dependences among them. In some cases, the structure can be generated through knowledge elicitation from domain experts. But usually the manual expert-based method is time-consuming because the entire structure should be defined by hand, and in many cases the suitable domain knowledge is not available (Corney, 2000). Another method is to learn Bayesian network structure from dada. But as a function of the number of nodes, the number of possible structures grows exponentially. So the exhaustive enumeration of all network structures is not feasible (Cooper & Herskovits, 1992). To solve this problem a number of search algorithms have been proposed. In this paper, the Greedy Thick Thinning algorithm (GTT) algorithm will be used to construct the Bayesian network for FF analysis. Generally the GTT algorithm works in two stages. Firstly, it starts with an empty graph and adds (or reverses) the arc (without creating a cycle) that increases the score as much as possible. This process is repeated until no adding or reversing can increase the score anymore. After that it repeatedly removes arcs until no arc deletion will result in a positive increase in the score (Dash & Druzdzel, 2003). Once the structure has been constructed, parameters (probabilities for each node) should be decided next. The easiest way of calculating the probabilities is using data frequency. However, as the size of data in the FF analysis study is usually not very large, using a frequency method will be not very suitable, as many feature combinations will not have been observed. To solve this problem, the Expectation Maximization (EM) algorithm will be used. The EM method can fill in the missing values with the most likely values as if the dataset is complete, and then it will iterate until convergence (Dempster, Laird, & Rubin, 1977). 3.3. FF analysis Now two Bayesian networks, one for capacity analysis and the other for complexity (lack of usability), have been constructed defining both the structure and conditional probabilities. We can use the Bayesian networks to conduct inference to the relationships among features and the whole product. To make inference we should firstly fix the states of known variables, and then propagate the beliefs around the network until all beliefs are consistent. After that we can get the desired probability distribution directly from the network. Belief updating is computationally complex, and in the worst case belief updating algorithms are NP-hard (Cooper, 1990). In this paper we use clustering algorithm proposed by Lauritzen and Spiegelhalter (1988) to deal with the updating task. This algorithm is fast and efficient to make belief updating in Bayesian network with hundreds of variables tractable. To find the key features which affect FF most, some concepts and techniques from information theory are used to deal with the analysis work. The concept of entropy in information theory was first introduced by Shannon (1948). Entropy is a measure of the uncertainty associated with a random variable. Let X be a discrete random variable, given value x, the probability distribution function is p(x) = Pr(X = x). The entropy H(X) of X can be defined as
HðXÞ ¼
X
pðxÞ log pðxÞ
ð2Þ
x2X
3.2. Bayesian network construction After getting the dataset of FF discussed in Section 3.1, the related Bayesian network will be constructed in this stage. Typically, there are two steps to build a Bayesian network (Heckerman,
Note that the entropy of X does not depend on the actual values of X, it only depends on p(x). Suppose X and Y are two random variables. Then for any fixed value y of Y, we can get a conditional probability distribution p(X|y) as follows:
10634
HðXjyÞ ¼
M. Li, L. Wang / Expert Systems with Applications 38 (2011) 10631–10637
X
pðxjyÞ logðpðxjyÞÞ
ð3Þ
x2X
Let the conditional entropy H(X|Y) be defined to be the expectation of H(X|y) when average over all possible values y. Then it is
HðXjYÞ ¼
XX
pðyÞpðxjyÞ logðpðxjyÞÞ
when using complex products, this criteria actually represents the comprehensive effect of FF in the view of companies. And the result of Fx minus Fa reflects the FF degree of the related product, and for FFD the lower the better. A more detailed analysis will be shown in the following example.
ð4Þ
y2Y x2X
4. Example case
Now to measure the mutual information of X and Y, I(X|Y) can be defined as
IðXjYÞ ¼ HðXÞ HðXjYÞ
ð5Þ
This measure represents the difference of the a priori and a posteriori entropies of X, i.e., the reduction in uncertainty about X by knowing Y. It is noted that I(X|Y) is the information in X about Y, and a high degree of dissimilarity indicates more information X carries about Y. As its name suggests, mutual information is symmetric, so I(X|Y) = I(Y|X). Entropy reduction is suitable for sensitivity analysis in Bayesian network (Lee, Park, & Shin, 2009), and in this paper we use it to identify the key features that affect FF most. After the key features have been selected, experts in the company can determine these features’ capacity and complexity level from both marketing and engineering concerns. Usually some candidate products will be given and their FF degree can be measured as
FFD ¼ Fx Fa
ð6Þ
where Fx is the complexity score, Fa is the capacity score. As Thompson et al. (2005) shown that, to defeat FF firms should balance the initial sales and long-term customer satisfaction. So in Eq. (6), FFD does not just reflect the fatigue feeling of consumers Table 1 Probabilities of the features in the capacity Bayesian network. Feature
HA
MA
LA
Digital camera Games GPS Wireless internet FM radio Email MP3 player Bluetooth
81.9 69.0 62.0 60.1 58.1 44.2 30.0 16.5
11.9 21.0 26.0 16.2 26.6 43.9 58.7 73.7
6.17 10.0 12.0 23.6 15.3 12.0 11.2 9.78
4.1. Case description To illustrate the approaches proposed in this paper, an example of FF analysis for smart phone development is presented in this section. Smart phones are typical high-feature products as manufacturers are lured into integration of a growing number of technologies and features to provide attractive and competitive models (Zheng & Ni, 2006). Customers initially choose these high-feature smart phones but after use they will find that the usability of such complex products becomes an increasing problem (Keijzers et al., 2008). In this paper eight major features of smart phone are selected to analyze their different effect on FF in the following paragraphs. 4.2. Capability and complexity measurement The selected features are listed in Table 1. And we also get 100 customers’ evaluations. As discussed in Section 3.1, the FF problem is directly related to two factors: capacity and complexity. To simulate the real market situation, the 100 ‘‘customers’’ are asked to respectively evaluate the capacity and complexity level of the eight features and the whole product. The capacity evaluation represents customer preferences at the point of purchasing, and after use the direct experience will affect the complexity level. A 3-level scale is used to present the evaluation: H (high), M (moderate) and L (low). For capacity the three levels are HA, MA and LA, and for complexity they are HX, MX and LX. 4.3. Bayesian network construction In this section two Bayesian networks are constructed based on the evaluation data. For each network GTT algorithm is firstly used to construct the structure and then EM algorithm is used for
Fig. 1. Bayesian network for capacity analysis.
10635
M. Li, L. Wang / Expert Systems with Applications 38 (2011) 10631–10637
Fig. 2. Bayesian network for complexity analysis.
parameter learning. The models described in this paper are created using the GeNIe modeling environment, which is a type of Bayesian network and decision support software and developed by the Decision Systems Laboratory of the University of Pittsburgh (http://dsl.sis.pitt.edu). The two networks for capacity and complexity evaluation are respectively shown in Figs. 1 and 2, in which FFA represent the capacity level of the whole product, and FFX means the product’s complexity level. 4.4. FF analysis for smart phone development Given the complete Bayesian network with both structure and conditional probabilities, we can make inferences and sensitivity analysis to guide the task of defeating FF. Firstly, Table 1 shows the probabilities of the features in Fig. 1 ordering by HA, which is the highest level of capacity. And Table 2 represents the probabilities in Fig. 2 ordering by the highest complexity level HX. Note that these probabilities result from the Bayesian network calculations rather than frequency data. From Eq. (6), we know a feature with low capacity level and high complexity level will have a high FF degree. Just from the result of Tables 1 and 2 the features like Bluetooth (HA = 16.5% and HX = 32.2%) should be paid attention to improve their capacity and/or reduce their complexity. But this simple conclusion ignores the relationships among features and their different effect on the whole product’s FF degree. More analysis will be made in the following paragraphs. As discussed in Section 3.3, entropy reflects the relative influence between uncertain variables. Tables 3 and 4 show the result of entropy reduction calculation from sensitivity analysis of
features related to FFA and FFX respectively. If set 0.05 as the threshold value for decision, the features influence FFA most are Wireless internet, games, MP3 player, Email and digital camera. And the features of MP3 player, digital camera and wireless internet have the greatest influence on FFX. The sensitivity analysis can help to identify the most influential decisions to defeat FF. For example, the Wireless internet feature has a much larger entropy reduction value than the feature of Bluetooth for both FFA and FFX. That means Wireless internet can provide more information and eliminate more uncertainty of FFA and FFX, while trying to confirm the Bluetooth feature will have less effect on the probability distribution of the whole product’s capacity and complexity.
Table 3 Entropy reduction values in the capacity Bayesian network. Feature
I(X; Y)
Wireless internet Games MP3 player Email Digital camera FM radio GPS Bluetooth
0.20756 0.15981 0.08770 0.07562 0.06231 0.02502 0.01056 0.00835
Table 4 Entropy reduction values in the complexity Bayesian network.
Table 2 Probabilities of the features in the complexity Bayesian network. Feature
HX
MX
LX
Feature
I(X; Y)
GPS Email Games Wireless internet MP3 player Bluetooth Digital camera FM radio
49.0 46.0 39.1 39.0 38.0 32.2 26.7 18.0
22.0 19.0 24.9 16.0 40.4 28.5 46.3 66.0
29.0 35.0 36.0 45.0 21.6 39.3 27.0 16.0
MP3 player Digital camera Wireless internet FM radio Email Bluetooth GPS Games
0.21741 0.16908 0.07478 0.01829 0.01698 0.01573 0.00359 0.00247
10636
M. Li, L. Wang / Expert Systems with Applications 38 (2011) 10631–10637
Table 5 FF evaluation results for candidate products. Figure
P1
P2
Capacity
Complexity
Capacity
Complexity
Digital camera MP3 player
HA = 100% MA = 100%
HX = 100% MX = 100%
MA = 100% HA = 100%
MX = 100% HX = 100%
Other features
Initial probability distribution
FFA FFX
HA = 41.5%
HA = 33.3% HX = 24.9%
HX = 23.3%
Based on the entropy reduction results, experts in the company can choose key features affect product’s FF most and decide their capacity and complexity level from both marketing and engineering concerns. Usually some candidates will be given in this step and the next task is how to evaluate them. In a Bayesian network, if the values of some variables are given, the probabilities of the remaining variables can be calculated. In this case, if one or some key features’ capacity and complexity level have been decided, the probability of FFA in Fig. 1 and FFX in Fig. 2 will be changed and can be calculated using the related network. An example of evaluation is shown in Table 5. P1 and P2 are two candidate products. For P1 the digital camera (DC) feature has a high capacity level HA and high complexity level HX. That means the DC feature is very attractive but hard to use. And the MP3 player of P1 is moderate for both capacity and complexity. Other features’ capability and complexity level is not decided and given the initial probability distribution. Using Figs. 1 and 2 the probability of high capacity level HA for P1 is 41.5%, and HX is 24.9%. For P2, its HA is 33.3% and HX is 23.3%. To evaluate these two products, Eq. (6) should be used. To be simple, we use the probability of HA to represent the capacity score Fa, and HX’s probability as the complexity score Fx. So the FF degree of P1 is 0.166 and P2 is 0.1. As P1 has a lower FF degree, it is better than P2 considering both capability and complexity. From Table 5 we can see even the probability of P1’s HX (24.9%) is a little larger than P2’s (23.3%), but P1 is much more attractive than P2 (41.5% versus 23.3%). Besides sensitivity analysis and evaluation presented above, Bayesian network can also be used to analyze the relationships among features, and this is useful in the process of deciding key features’ capacity and complexity level. Figs. 3 and 4 show the effect of setting the value of MP3. In Fig. 3 the first bar of each pair represents the prior probability of the variable having a ‘‘HA’’ value, and the second bar is the corresponding posterior probability after the value of some variables like MP3 has been decided. And Fig. 4 shows the probability change of HX. We can find that when the value of MP3 is ‘‘HA’’, the probability of FM’s HA value has increased from 0.581 to 0.721. This increase indicates that the capacity of MP3 and FM are correlated to some extent, that is, a HA MP3
Fig. 4. Prior and posterior probabilities of HX when set MP3 to HX.
tends to ‘‘cause’’ the increase of FM’s HA level probability. We can also find that the FFA increase from 0.511 to 0.753. That means a high capacity MP3 will also bring a significant capacity enhancement of the product. The same process can also be used for the analysis of complexity relationships among different features in Fig. 4. 5. Conclusion In this paper, a probability based methodology was proposed to analyze the effect of FF on company’s long-term revenue. Bayesian network techniques were used to represent the uncertain customer preferences for capacity and usability. In order to construct a Bayesian network, customers were asked to evaluate the capacity and complexity (lack of usability) level of the whole product and each individual feature. Based on these evaluation data Bayesian networks were built and used for the FF analysis task. Sensitivity analysis was implemented based on entropy reduction to identify the key features that affect product’s FF most. Experts in the company could determine the capacity and complexity level of these key features from both marketing and engineering concerns. After that some candidate products were constructed, and their FF degree could be got through calculation based on the Bayesian networks. And using inference techniques the relationships among features were analyzed by comparing the prior and posterior probabilities. The method proposed in this paper can guide decision-makers in marketing and engineering to make the most influential decisions in the process of product development. As the application of probabilistic graphical models, the uncertainty problem of customer preferences for capability and complexity could be handled efficiently. If enough data are hard to get in some situations, the Bayesian network can be constructed through experts’ domain knowledge and the proposed method can still work for FF analysis. As a closing remark, we are aware of the limitations within this study. One is that to build a Bayesian network, the features contained in a product should be given first. If decision-makers want to delete/add one or some features, they should rebuild Bayesian network from the initial data gathering step. Another limitation is that the data used in this study just focuses on ‘‘present’’ customer preferences. In fact, customer preferences are dynamic and may vary drastically from time to time (Chen & Wang, 2008). So our future work is to be aimed at dealing with the problem of changeable features and dynamic customer preferences in the process of defeating FF. Acknowledgements
Fig. 3. Prior and posterior probabilities of HA when set MP3 to HA.
This research is supported by the National Natural Science Foundation of China (Grant No. 70932004/G0209), and the
M. Li, L. Wang / Expert Systems with Applications 38 (2011) 10631–10637
National Natural Science Foundation of China (Grant No. 71072061/G020801). References Chen, Z. X., & Wang, L. Y. (2007). Integrating knowledge acquisition and deduction in product configuration by neural networks. In The third IEEE international conference on wireless communications, networking and mobile computing. Chen, C., & Wang, L. (2008). Integrating rough set clustering and grey model to analyse dynamic customer requirements. Proceedings of the Institution of Mechanical Engineers, Part B: Journal of Engineering Manufacture, 222(2), 319–332. Cooper, G. F. (1990). Research note the computational complexity of probabilistic inference using Bayesian belief networks. Artificial Intelligence, 42, 393–405. Cooper, G. F., & Herskovits, E. (1992). A Bayesian method for the induction of probabilistic networks from data. Machine Learning, 9(4), 309–347. Corney, D. (2000). Designing food with Bayesian belief networks. Adaptive computing in design and manufacture. Univ. of Plymouth. pp. 83–94. Dash, D., & Druzdzel, M. J. (2003). Robust independence testing for constraint-based learning of causal structure. Proceedings of the nineteenth conference on uncertainty in artificial intelligence, 167–174. Dempster, A. P., Laird, N. M., & Rubin, D. B. (1977). Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical Society. Series B (Methodological), 39(1), 1–38. Ellison, G. D. (2005). A model of add-on pricing. The Quarterly Journal of Economics, 120(2), 585–637. Gill, T. (2008). Convergent products: What functionalities add more value to the base? Journal of Marketing, 72(2), 46–62. Green, P. E., & Krieger, A. M. (1991). Segmenting markets with conjoint analysis. The Journal of Marketing, 20–31. Hamilton, R. W., & Thompson, D. V. (2007). Is there a substitute for direct experience? Comparing consumers’ preferences after direct and indirect product experiences. Journal of Consumer Research, 34(4), 546–555. Hauser, J. R., & Rao, V. R. (2003). Conjoint analysis, related modeling, and applications. Marketing Research and Modeling: Progress and Prospects: A Tribute to Paul Green, 141–168. Heckerman, D. (1996). A tutorial on learning Bayesian networks. Technical Report MSRTR-95-06, Microsoft Research. Heckerman, D. (1997). Bayesian networks for data mining. Data Mining and Knowledge Discovery, 1(1), 79–119. Jiao, J., Zhang, Y., & Helander, M. (2006). A Kansei mining system for affective design. Expert Systems with Applications, 30(4), 658–673.
10637
Keijzers, J., den Ouden, E., & Lu, Y. (2008). Usability benchmark study of commercially available smart phones: Cell phone type platform, PDA type platform and PC type platform. In Proceedings of the 10th international conference on Human computer interaction with mobile devices and services (pp. 265–272). Kruger, J., Galak, J., & Burrus, J. (2007). When consumers’ self-image motives fail. Journal of Consumer Psychology, 17(4), 250–253. Lauritzen, S. L., & Spiegelhalter, D. J. (1988). Local computations with probabilities on graphical structures and their application to expert systems. Journal of the Royal Statistical Society. Series B (Methodological), 50(2), 157–224. Lee, E., Park, Y., & Shin, J. G. (2009). Large engineering project risk management using a Bayesian belief network. Expert Systems with Applications, 36(3P2), 5880–5887. Löfgren, M., & Witell, L. (2008). Two decades of using Kano’s theory of attractive quality: A literature review. The Quality Management Journal, 15(1), 59. Lu, J., Bai, C., & Zhang, G. (2009). Cost-benefit factor analysis in e-services using Bayesian networks. Expert Systems with Applications, 36(3P1), 4617–4625. Mandel, N., & Nowlis, S. M. (2008). The effect of making a prediction about the outcome of a consumption experience on the enjoyment of that experience. Journal of Consumer Research, 35(1), 9–20. Nowlis, S. M., & Simonson, I. (1996). The effect of new product features on brand choice. Journal of Marketing Research, 33(February), 36–46. Rahman, M., & Rahman, M. M. (2009). To defeat feature fatigue the right way, understand it first. Strategic Direction, 25(6), 26–28. Rust, R. T., Thompson, D. V., & Hamilton, R. W. (2006). Defeating feature fatigue. Harvard Business Review, 84(2), 98. Shannon, C. E. (1948). A mathematical theory of communication Reprinted with corrections from. The Bell System Technical Journal, 27(379–423), 623–656. Tep, S.-B. S. P. (2009). Web site affective quality and features quantity in relation to customer online satisfaction: A research framework. In The 3rd student conference on business research. Thompson, D. V., Hamilton, R. W., & Rust, R. T. (2005). Feature fatigue: When product capabilities become too much of a good thing. Journal of Marketing Research, 42(4), 431–442. Tsai, H. M., & Liu, C.H. (2007). The influence of feature fatigue on the preferences of hedonic and utilitarian features. In International conference on business and information. Venkatesh, R., & Mahajan, V. (1993). A probabilistic approach to pricing a bundle of products or services. Journal of Marketing Research, 30(November), 494–508. Wang, Y. & Tseng, M. M. (2007). An approach to improve the efficiency of configurators. In 2007 IEEE international conference on industrial engineering and engineering management (pp. 1332–1336). Zheng, P., & Ni, L. M. (2006). Smart phone and next generation mobile computing. San Francisco: Morgan Kaufmann.