Recommendation in Social Media: Utilizing Relationships among Users to Enhance Personalized Recommendation

Recommender systems are ubiquitous in our digital life in recent years. They play a significant role in numerous Internet services and applications such as electronic commerce (Amazon and eBay), on-demand video streaming (Netflix and Hulu). A key task in recommender systems is to model user preferences and to suggest, for each user, a personalized list of items that the user has not experienced, but are deemed highly relevant to her. Many of these recommendation algorithms are based on the principle of collaborative filtering, suggesting items that similar users have consumed. With the advent of online social networks, social recommendation has become one of the most popular research topics in recommender systems, exploiting the effects of social influence and selection in social networks, where user relationships are explicit, i.e., there will be an edge connecting two users if they are friends. In addition, more information about the relationships between users in social media becomes available with the rapid development of various Internet services. For example, more and more online web services are providing mechanisms by which users can self-organize into groups with other users having similar opinions or interests, enabling us to analyze the interactions between users with others insides/outsides groups, as well as the engagement between users and groups. User relationships in these applications are usually implicit and can only be utilized indirectly for recommendation tasks. In this thesis, we focus on utilizing user relationships (either explicit or implicit) to enhance personalized recommendation in social media. We study three problems of recommendation in social media, i.e., recommendation with strong and weak ties, social group recommendation and interactive social recommendation in an online setting. We propose to improve social recommendation by incorporating the concept of strong and weak ties which are two well documented terms in the social sciences, boost the performance of social group recommendation through modeling the temporal dynamics of engagement of users with groups, and tackle the interactive social recommendation problem via employing the exploitation-exploration strategy in an online setting. Our proposed models are all compared with state-of-the-art baselines on several real-world datasets.