Personalized recommender systems on the social web

An exponential increase in the usage of the World Wide Web (Web 2.0) has led to a wide collection of social platforms that harness people's online activities such as online shopping (e.g., eBay and Amazon), entertainment (e.g., YouTube and Last.fm), and social networking (e.g., Facebook and Twitter). This has led to an overabundance of information online, which is far too consuming for individual users to browse. As such, personalized recommender systems have emerged, which aim to tailor the online information presented to the user based on their preferences. Such recommender systems have been widely studied in the research community and are also commonly deployed in industry, e.g., music recommendation on Last.fm, item recommendation on Amazon.com, and friend recommendation on Twitter. Despite the huge prevalence of using social annotations in personalized recommendation, a major outstanding problem is that in general they are redundant, sparse and ambiguous. Therefore, in this thesis, we first study and address this problem by proposing a recommendation-oriented deep neural model, in the context of tag-aware personalized recommendation. Here, we also propose hybrid deep learning and negative sampling to enhance the deep models' training efficiency. Furthermore, since users' real-time news preferences are sometimes correlated with their geographical context, we then investigate location-aware personalized news recommendation using geographical topic models, where topic feature modelling and the user's long-term personal interests are used to enhance the recommendation performance. Finally, to alleviate the need for powerful computing facilities to train deep recommendation models, we propose a novel ontology-based similarity to circumvent tag ambiguity and redundancy for applications in light-weight recommender systems on the Social Web.