Insights from Applying Sequential Pattern Mining to E-commerce Click Stream Data

Previous sequential pattern mining algorithms have focused on improving performance in terms of runtime and memory consumption without considering the specifics of different data sources or application scenarios. In this paper, we focus on mining closed sequential patterns from website click streams by extending the state of the art Bi-Directional Extension (BIDE) algorithm in order to identify domain-specific rule sets. In particular, we focus on exploiting sequential patterns for landing page personalization and product recommendation in the e-commerce domain. Our contribution is therefore of algorithmic as well as of empirical nature. Based on a dataset that we derived from an online store for nutritional supplements, we evaluate the effectiveness of using different sources of domain knowledge, such as product hierarchies and search word categorizations, to enhance predictions about the conversion actions of users. Furthermore, we examine the performance of the recommender for two important user subgroups, namely those that use search functionality and those that don't. Our findings indicate for instance that search terms alone are already quite effective for predicting users' add-to-basket actions and that using additional domain knowledge to generate multi-dimensional rules does not always lead to improved accuracy.