Learning item Temporal dynamics for predicting buying sessions

Identifying whether an e-commerce session will end up in a buy is an ongoing research topic. A session predicted as a non-buying one may trigger recommender systems, thus increasing the probability of a buy. Alternatively, a session predicted as a buying session may enable recommender systems to predict additional items. In this work, we suggest a prediction model leveraging the temporal characteristics of both the session and the items clicked in that session. Our method introduces a buying probability per session as a function of the clicked-items recent purchase history, and the session temporal characteristics. Empirical results on imbalanced e-commerce dataset with more than nine million sessions demonstrate that we achieve high Precision, Recall and ROC in predicting whether a session ends up with a purchase. In a wider perspective, our findings shed light on the importance of considering items temporal dynamics in e-commerce sites recommendations.