Effects of Session Representation Models on the Performance of Web Recommender Systems

Abstract

Usage pattern discovery is a crucial part of Web recommendation systems and is strongly related with the performance of the recommender. User sessions which are constructed from Web access logs give information about a user's navigational behavior. These session records are one of the main resources in the pattern discovery stage. The representation of the sessions can be in different forms such as feature vectors based on page durations or existence/nonexistence of pages. This work analyzes the effects of the different recommendation models which consider different characteristics of user sessions. For this purpose, we used three different recommender models. The first one considers only the existence of the visited pages in a session and the view time of each page. The second recommender model considers only the order of the visited pages in each session. The third model is based on the co-occurrence of the visited pages among user sessions. Our experimental results show that using the ordering information improves the prediction accuracy of the next request.