Adaptive Bayesian Non-Parametric user profile classification in e-commerce websites

On e-commerce sites, users change their pref-erences over time concerning the products they want to purchase. Indeed, a product that is desired today may no longer be of interest tomorrow. This may be caused by a purchase of a similar product, a change of user's situation (marriage, birth, etc.) leading to the appearance of new preferences. Therefore, classifying users of an e-commerce site in a fixed class can induce inaccurate recommendations. Besides, a user may have different tastes and behaviors on the e-commerce site and therefore belong to several classes of user profiles. Finally, the number of classes of user profiles cannot be known a priori and may change over time since the membership of new users to the site may include the appearance of new classes. To overcome these issues, state-of-the-art approaches are mostly based on sequential models to keep a trade-off between the historical interests and the current session interests. However, these models fail to represent users' relation and discover new trends in user profiles. In this paper, we handle these issues and propose a model-based Bayesian non-parametrics. Our proposed solution can select an adaptive number of users' profiles and point out the discriminating hidden interests of users. Furthermore, the method could be applied to any e-commerce dataset and does not rely on feature engineering or specific parameters. Experiments are performed using real data from publicly available data sets, and the obtained results demonstrate the adaptability of the approach and its ability to infer the hidden features behind the appearance of user profiles.