User Profiling with Hierarchical Context: An e-Retailer Case Study

Abstract

In e-commerce applications, no systematic research has been provided to evaluate if the use of a detailed and rich contextual representation improves the user modeling predictive performances. An underestimated issue is also evaluating if context could be inferred by existing customer data off-line, in spite of getting the customer involved on-line in the gathering process. In this paper, we address those problems, defining context as "the intent of" a customer purchase. To this aim, we collected data containing rich contextual information, hierarchically structured, by developing a special-purpose browser. The experimental results show that the finer the granularity of contextual information the better is the modeling of customers' behavior. Representing the context in a hierarchical structure is a necessary condition, for inferring the context off-line, but it's not a sufficient one.

A hybrid approach for predicting customers’ individual purchase behavior

“…These findings indicate that the individual-level and the segment-based approaches complement each other really well, and the hybrid approach based on these techniques is able to overcome the shortcomings of these state-of-the-art approaches. Furthermore, the comparative results between the individual-level and the segment-based approaches corroborate the findings of the previous studies (Jiang and Tuzhilin, 2006; Lombardi et al , 2013; Palmisano et al , 2007, 2008) which have shown that the individual-level technique provides better prediction performance than the segment-based technique. The comparative results also show that among the five algorithms, the logistic regression achieves the highest prediction performance.…”

“…Jiang and Tuzhilin (2006, 2009) used these approaches to predict customer purchase behaviors in both online and “brick-and-mortar” stores, and their findings suggest that while the individual-level customer modeling approach dominates the segment-based approach for high-transaction customers, segment-based approach outperforms the individual-level approach for low-volume customers. Palmisano et al (2007, 2008) incorporated contextual information into customer behavior models and conducted performance comparisons of the individual-level and the segment-based approaches in building of these models. Their findings demonstrate that the individual-level approach leads to higher predictive performance for pure models, whereas segment-based approach tends to perform better than the individual-level approach when the contextual information of the transactions is included to customer models.…”

Using Context to Improve Predictive Modeling of Customers in Personalization Applications

“…Also, the question of how to represent context, once it is inferred to model customers' behavior, has not been addressed before, except in [1] where it was only tangentially referred to and explored within recommender systems. This paper is built on our previous work on contextual profiling [14], [24]. In particular, we have consolidated and enhanced our prior studies of questions 1 and 2 from [14] and [24], respectively.…”

“…This paper is built on our previous work on contextual profiling [14], [24]. In particular, we have consolidated and enhanced our prior studies of questions 1 and 2 from [14] and [24], respectively. We also added question 3, expanded our data analysis to an additional data set, and integrated studies of questions 1 and 3 into a self-contained journal paper.…”

“…This true story is very symptomatic of problems pertaining to many personalization systems that often infer customer behavior from the registration and the purchasing information of online customers without studying the contexts in which these purchases are made. In the previous example, if the system knew that the purchase was made in the context of a gift, this transaction should have been discarded from inferring that customer's behavior, and the whole problem would have been avoided [14], [24]. Getting such contextual data, such as the intent of a purchase, is not easy in many applications.…”