Constructing a Context-aware Recommender System with Web Sessions

Abstract

During the last decade, the importance of recommender systems has been increasing to the point that the success of many well-known service providers depends on these technologies. Recommender systems can assist people in their decision making process by anticipating preferences. However, common recommender algorithms often suffer from lack of explicit feedback and the \cold start" problem. This thesis investigates an approach of using implicit data only, to extract users' intent for fashion e-commerce in cold start situations. Markov Decision Processes (MDPs) are used on web session data to extract topic models. This thesis also explores how well the topic models can capture users intent and whether they can be used to produce good recommendations. The results show that this approach was able to accurately identify sessions topics, and in most cases the topics could successfully be translated to product recommendations.