Improve serendipity in movies recommendation usings random walks with restart

Details

Supervisors

Saerens, Marco

Faculty

Louvain School of Management

Degree label

Master [120] en Ingénieur de gestion

Abstract

The goal of this master thesis is to improve the serendipity in the recommendation of movies. Serendipity is a concept that gathers three notions: relevancy, novelty and unexpectedness. Serendipity item is a useful and surprising item. The literature has showed that classic recommender systems recommend products already known by the users of the recommender system. Moreover, classic recommender systems recommend items in the same genre of the past acquisitions of the user. If a user bought a physics book, recommender system will suggest him other physics book. It is not pertinent because if you already have a physics book, you do not want a second. For all these reasons, the building of recommender system that improves serendipity is a major goal in machine learning. The thesis is focused on movies recommendation and uses a database that contains rating of items by users. This database is MovieLens. In this thesis, the model used to recommend items is the random walk with restart. Random walk with restart is a known mathematical model used for instance to rank webpages when you type a query on a search engine. Random walk with restart has already used in recommender system, but in this thesis a bias is implemented in the random walk to give more importance to less known movies. The goal is to avoid recommending blockbuster and to recommend less known movies while keeping relevant recommendations. The bias is made according to a method applied in stochastic routing of packages in scale-free networks (Fronczak and Fronczak, 2009). This type of bias was never applied to bias a random walk with restart. The results of this model were compared to a classic model of recommendation. The results of the three criteria relevancy, novelty and unexpectedness were computed for the simple random walk with restart, the biased random walk with restart and the classic model. Finally, the conclusion summarize if this technique is effective or not according to the serendipity concept.