Recommendation technologies have become increasingly important due to the ubiquity of information overload across various application domains including E-commerce(e.g., Amazon), online entertainment (e.g., Netflix, Pandora) and publishing (e.g., Google News). As the technology and application of recommendation is rapidly evolving in these years, traditional collaborative filtering algorithms such as nearest neighbour or matrix factorization have fallen short in coping with several emerging but critical issues in modern applications. Firstly, ranking items, especially identifying a few most interesting items out of a huge pool, has become the core task in most application scenarios. In this thesis, we propose a new class of algorithms for directly solving the personalized ranking problem...[ Read more ]

Recommendation technologies have become increasingly important due to the ubiquity of information overload across various application domains including E-commerce(e.g., Amazon), online entertainment (e.g., Netflix, Pandora) and publishing (e.g., Google News). As the technology and application of recommendation is rapidly evolving in these years, traditional collaborative filtering algorithms such as nearest neighbour or matrix factorization have fallen short in coping with several emerging but critical issues in modern applications. Firstly, ranking items, especially identifying a few most interesting items out of a huge pool, has become the core task in most application scenarios. In this thesis, we propose a new class of algorithms for directly solving the personalized ranking problem by representing user feedback using pairwise preference based representation. Secondly, we extend the proposed ranking model to also consider the temporal context, as time awareness is becoming an increasingly important feature in real world applications, which often need to cope rich temporal dynamics and provide context aware recommendations. Finally, we further the extend the framework to also consider relational information about users and/or items. In particular, we consider the social relations among users the taxonomical relations between items, which are commonly found in real world systems. Our results demonstrate that utilizing these contextual information could greatly improve upon context-oblivious algorithms under data sparsity conditions.