Recommendation with contextual information

Information retrieval (IR) systems have tremendously broaden users' access to information. However, users need to select their needs from trillions of information indexed daily. Due to the "semantic gap" between queries and indexed terms in IR system, whether users can satisfy their needs depends on whether they use the correct terms as queries. Recommender systems, have become a counterpart of information retrieval systems such as Google. They do not require users to specify their information needs in advance. They model users' preferences based on their history data, and automatically recommend items which satisfy their needs. In this way, both semantic gap and information overload can be alleviated. Collaborative filtering is the most popular recommendation algorithm used in academia and industry. However, it suffers from the cold start problem, where it cannot recommend new items to users or give recommendation to new users. Its recommendations also bias towards popular items.The goal of this thesis is to develop recommendation algorithms which can solve the item-side and user-side cold-start problems. We incorporate item content to solve the item-side cold start problem, and incorporate user similarity networks or social networks to solve the user-side cold-start problem. Our contributions are listed as follows. First, we develop a content-based movie recommendation algorithm which incorporates user, actor, and movie into the same model. Second, we develop a content-based music recommendation algorithm. The core component is a hierarchical dynamic programming algorithm to compute music similarities. The resulting music similarity matrix is then employed to make recommendations. Third, we apply social recommendation algorithms on Douban dataset. We look into a "Twitter-like" user-follower social network to see if it can improve recommendation performance. Finally, we validate the existence of social influence. We also analyze the social influence by investigating the correlation between one's influence and centrality, and by investigating how recommendation effectiveness changes with the number of recommendations.Ph.D., Information Studies -- Drexel University, 201