Semantics-aware user modeling and recommender systems in online social networks

Abstract

The popularity of Online Social Networks (OSNs) has rapidly increased over the past few years. User modeling (including creating user interest profiles) and recommendation approaches in OSNs are important methods to deal with cold-start problems inside or outside those OSNs and to cope with the problem of information overload. Recently, semantics-aware techniques such as top-down approaches which explore external knowledge sources such as DBpedia and bottom-up approaches which learn latent semantic representations for users/items using factorization or embedding approaches have received great attention in the domain of user modeling and recommender systems.

The aims of this study were to propose semantics-aware approaches for user modeling and recommending items in OSNs. For active users who consistently generate content, various user modeling dimensions such as the temporal dynamics and semantics of user interests, and a comprehensive user modeling strategy considering those dimensions, have been investigated. For passive users who only follow other users in OSNs without generating content, various types of information about their followees (the users they follow in OSNs) have been investigated. Furthermore, this thesis investigates semantics-aware recommendation approaches based on semantic information from knowledge graphs (KGs) such as DBpedia, and proposes semantic similarity/distance measures and factorization approaches in the context of different recommendation scenarios such as a cold start.

The experimental results show that the strategy for representing user interests plays the most important role followed by the temporal dynamics of user interests in user modeling for active users. For passive users, the results show that both biographies and list memberships of followees provide useful information for inferring user interest profiles, and the profiles inferred based on this information outperform the ones inferred based on the information from the tweets or account names of followees. mLDSD, a proposed semantic similarity measure with a global normalization strategy outperforms other semantic similarity measures in the context of a cold-start scenario for item recommendations. When there is plenty of feedback from users, LODFM, a proposed factorization approach exploring lightweight DBpedia features outperforms other state-of-the-art methods significantly in two different domains. As the incompleteness of KGs had not been considered for semantics-aware recommendations in the literature, we further investigated transfer learning between item recommendations and knowledge graph completion. The results showed that considering the incompleteness of a KG can further improve the performance when compared to LODFM, and performs better than other baselines. In addition, the results show that exploiting user-item interaction histories also improves the performance of completing the KG with regard to the domain of items, which has not been investigated before.