The effect of context on misclassification costs in e-commerce applications S. Lombardi a , M. Gorgoglione b , U. Panniello b,⇑ a Iveco S.p.A, 10156 Torino, Italy b Politecnico di Bari, Department of Mechanical and Business Engineering, Viale Japigia 182, 70126 Bari, Italy article info Keywords: Data mining Business intelligence Knowledge management applications Context Predictive model abstract The performance of customer behavior models depends on both the predictive accuracy and the cost of incorrect predictions. Previous research showed that including context in the customer behavior models can improve the accuracy. However, improving accuracy does not necessarily mean that the misclassifi- cation cost decreases. In fact, different errors have different costs. Even if the number of incorrect predic- tions decreases, the number of errors associated with higher costs increase. The aim of this paper is to understand whether including context in a predictive model reduces the misclassification costs and in which conditions this happens. Experimental analyses were done by varying the market granularity, the dependent variable and the context granularity. The results show that context leads to a decrease in the misclassification cost when the unit of analysis is the single customer or the micro-segment. The exceptions may occur when the unit of analysis is a segment. These findings have significant impli- cations for companies that have to decide whether to gather context and how to exploit it best when they build predictive models of the behavior of their customers. Ó 2013 Elsevier Ltd. All rights reserved. 1. Introduction In today’s intensely competitive, rapidly changing and highly complex business environment, the customer’s actions are less and less referable to rigid schemes. The variability of customers’ decision-making process, in fact, decreases the capacity to predict their behavior. It is important, therefore, to build customer behav- ior models which are more sophisticated and able to reveal homo- geneous patterns. Improved predictive modeling can lead to more effective marketing actions, efficient processes, higher levels of customer satisfaction, and higher returns on investments. Electronic commerce can be viewed as an ideal domain of application for data modeling technology (Provost, Melville, & Saar-Tsechansky, 2007). Clickstream data records the user’s activ- ity and contains the ‘‘trajectory’’ of clients (Van Den Poel & Buckinx, 2005). Processing this data offers the opportunity to im- prove the understanding of customer activities and to create a user profile that enhances the prediction of choice behavior. However, customers’ behavior depends on the context in which a transaction takes place. Several scholars have maintained that a change in the context (such as the intent of an Internet search, special payment conditions, or the user’s geographic location) makes the behavior of a customer change (Bettman, Luce, & Payne, 1998; Lussier & Olshavsky, 1979). Only very recently an experimental research has proved that including contextual information in a customer’s behavior model can increase the ability of predicting the cus- tomer’s behavior and, in turn, personalizing the marketing actions (Gorgoglione, Palmisano, & Tuzhilin, 2006; Palmisano, Tuzhilin, & Gorgoglione, 2008). However, the performance of user behavior models depends on both the predictive accuracy and the cost of incorrect predictions. Improving the accuracy does not necessarily mean that either the cost of wrong predictions decreases or the benefit overcomes the cost. For instance, the cost of mailing to a non-respondent is very small, but the cost of not mailing to someone who would respond is the entire profit lost. Moreover, the cost of false positives and false negatives greatly varies depending on the application goal and specific industry. For example, in the fraud detection applications, if the goal is to predict whether a card is false or not, the cost of a false negative is higher than the cost of a false positive because the false positive cost (false alarms) is only represented by a phone call to the cardholder ask- ing him to verify a couple of purchases. The cardholders do not re- sent these phone calls (as long as they are infrequent) so the cost is just a few minutes of operator time. Therefore, in this case a good predictive model should be primarily able to reduce the number of false negatives because the cost associated with them is higher. In fact, it corresponds to letting an actual fraud go undetected. 0957-4174/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.03.009 ⇑ Corresponding author. Tel.: +39 080 5962765; fax: +39 080 5962788. E-mail addresses: sabrina.lombardi@iveco.com (S. Lombardi), m.gorgoglione@ poliba.it (M. Gorgoglione), u.panniello@poliba.it (U. Panniello). Expert Systems with Applications 40 (2013) 5219–5227 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa