استفاده از زمینه برای بهبود اثربخشی تقسیم بندی و هدف گیری در تجارت الکترونیکی

In e-commerce, where competition is tough and customers’ preferences can change quickly, it is crucial for companies to segment customers and target marketing actions effectively. The process of segmentation and targeting is effective if the customers grouped into the same segment show the same behavior and reaction to marketing campaigns. However, the link between segmentation and targeting is often missing. Some research contributions have recently addressed this issue, by proposing approaches to build customer behavior models in each segment. However customers’ behavior can change with the context, such as in many e-commerce business applications. In these cases, building contextual models of behavior would provide better predictive performance and, in turn, better targeting. However, the problem of including context in a segmentation model and building predictive behavior model of each segment consistently is still an open issue. This research aims at providing an answer to the following research issue: how to include context in a segmentation model in order to build an effective predictive model of customer behavior of each segment. To this aim we identified three different approaches and compared them by a set of experiments across several settings. The first result is that one of the three approaches dominates the others in certain conditions in our experiments. Another important result is that the most accurate approach is not always the most efficient from a managerial perspective.

Today’s marketing professionals are being pressured to keep their companies competitive as significant changes are taking place in the business environment. On the one hand globalization, technological advances in information and communication technology (ICT), increasing competition, fragmentation of markets, have compelled companies to rethink their marketing strategies and processes. This is especially true on the web, where teradata of information are available every day and the search costs are low, thus making competition just “a mouse click away” (Jiang & Tuzhilin, 2009b). Therefore, in such increasingly competitive environment, it is crucial to segment customers and target marketing actions effectively. The process of segmentation and targeting is effective if the customers grouped into the same segment show the same behavior and reaction to marketing campaigns. E-commerce provides companies with the unprecedented opportunity to record data on customers behavior in richer way, thus making it possible to adopt better segmentation models and targeting marketing actions more effectively. Several studies have addressed this issue by using different techniques (Chan, 2008, Hwang et al., 2004, Jonker et al., 2004 and Kim et al., 2006). Some scholars have proposed to build predictive models of customer behavior in each segment in order to enable an effective targeting (Apte et al., 2001, Jiang and Tuzhilin, 2009a and Jiang and Tuzhilin, 2009b). Customers’ behavior can change with the context and keeping segmentation and targeting linked can remain challenging. In these cases, building contextual models of behavior would provide better predictive performance and, in turn, better targeting. The marketing literature (Bettman et al., 1998 and Lilien et al., 1992) has recognized the importance of considering contextual information because it can induce important changes in customer purchasing behavior, thus challenging traditional approaches to segmentation (Firat and Shultz, 1997, Peltier and Schribrowsky, 1992 and Tsai and Chiu, 2004). Research has shown that including the context in which a transaction occurs in customer behavior models, improves the ability of predicting their behavior (Gorgoglione et al., 2006 and Palmisano et al., 2008). However, no systematic study has been done before to empirically address the issue of how to incorporate context into segmentation and targeting and it still is an open issue. In this paper, we contribute to filling this gap. In particular, we present a conceptual framework to incorporating context into a segmentation model and building predictive models of customer behavior in each segment to improve targeting. We identified three possible approaches, namely contextual pre-filtering, contextual post-filtering and contextual profiling, and performed a comparative analysis across a wide range of experimental conditions. Each approach differs from the others in how context is considered along the entire process of building segment-based behavior models. Each type of contextual model is compared with an un-contextual segmentation, which does not take into consideration any contextual information. Some managerial implications arising when using different approaches are discussed.

In this paper, we presented a conceptual framework to incorporating context in a segmentation model. We identified three possible approaches, namely pre-filtering, post-filtering and profiling, and compared them by measuring the predictive accuracy of behavior models built in each segments. To this aim, we compared each one of the approaches to the un-contextual segmentation (e.g., segmenting customers without taking into account context). We conducted the comparison by a set of experiments across several settings: different data sets, predictive models, dependent variables, performance measures and segmentation techniques. We discussed how data sparsity and homogeneity affect the performance of segmentation of each approach and finally draw some managerial implications. The main results coming from the experiments can be summarized as follows. The contextual pre-filtering approach outperforms the un-contextual and the results are statistically significant. This effect increases with both market granularity (i.e., from macro to micro segments), and contextual degree of knowledge (i.e., from a rough to finer contextual information). On the contrary, the contextual profiling approach is either equivalent to the un-contextual (meaning that the differences in performance are not statistically significant) or it underperforms the un-contextual (when the difference are significant). One of the reasons is that this approach entails generating new segmentation variables and the increased number of attributes in customers’ profiles decreases the performance of the predictive model of behavior built in each segment. The contextual post-filtering approach outperforms the un-contextual in some experimental conditions. The reason is that introducing context later in the whole segmentation and targeting process means to decrease the effect of homogeneity. Therefore, only when context strongly affects customer behavior the post-filtering contextual approach provides better performance. From the managerial perspective, using contextual pre-filtering approach raises a practical concern related to the cost of a segmentation model. In fact, pre-filtering can generate many segmentation solutions thus making the costs increase. Therefore, companies should consider the use of a post-filtering approach depending on the business conditions because it can represent a good balance between predictive accuracy (better targeting) and cost. The choice should be made by carefully considering benefits and costs. Nevertheless, our experiments suggest a general rule of thumb. When context strongly affects customer behavior, thus making it possible to clearly recognize different models of behavior in different contexts, a post-filtering approach is likely to represent the best choice. When the effect of context is less strong, then it is better to introduce it earlier in the segmentation and targeting process: in these cases, the pre-filtering approach is likely to represent the best choice, although more expensive to manage. As an extension to our research, we plan to compare the performance of the contextual pre-filtering approach built at segment level, with individual customer models, in order to understand whether incorporating context at “the best way”, i.e. using contextual pre-filtering, would give better performance for segment-based models than individual models.