Purchase Prediction and Item Suggestion based on HTTP sessions in absence of User Information

In this paper, the task is to determine whether an HTTP session buys an item, or not, and if so, which items will be purchased. An HTTP session is a series of item clicks. A session has type buy, if it buys at least one item, or non-buy otherwise. Accordingly, data is in (session, item, time) format, which tells us when an item is clicked or purchased during an HTTP session. The main challenge comes from the fact that (1) user information is not available for clicked or purchased items, which are merely tagged with anonymous sessions, and (2) suggestions are highly temporal as they are suggested to sessions instead of users. In other words, users which are stable and identified are replaced with sessions which are temporal and anonymous. In this work, we propose a feature-based system that predicts the type of a session, and determines which items are going to be purchased. As the main contribution, we have modeled sessions separated by the number of unique items, prioritized item-features based on the number of clicks, and utilized cumulative statistics of similar items to attenuate the sparsity problem.