A long-standing goal of recommendation system is to solve data sparsity issue, where user-item preference data is not enough to train a reliable model. Significant strides have been made towards to this goal over the last few years thanks to the fast-moving fields of data gathering, algorithms and computing infrastructure. The progress has been especially rapid in offering context information going beyond user-item preferences. The context can be divided into two categories in terms of its resources type: in-domain context regarding attribute information associated with users and items, and cross-domain context concerning data from different but related domain. By developing methods that can effectively utilize context information for recommendation task, performance can be enhanced fu...[ Read more ]

A long-standing goal of recommendation system is to solve data sparsity issue, where user-item preference data is not enough to train a reliable model. Significant strides have been made towards to this goal over the last few years thanks to the fast-moving fields of data gathering, algorithms and computing infrastructure. The progress has been especially rapid in offering context information going beyond user-item preferences. The context can be divided into two categories in terms of its resources type: in-domain context regarding attribute information associated with users and items, and cross-domain context concerning data from different but related domain. By developing methods that can effectively utilize context information for recommendation task, performance can be enhanced further.

In the first part of this dissertation, we consider the problem of improving recommendation with in-domain context data. In movie industry, it is often the case that movie posters and stills deliver rich knowledge for understanding movies as well as users’ preferences. For instance, user may want to watch a movie at the minute when she/he finds some released posters or stills attractive. Unfortunately, such unique features cannot be revealed from rating data or other forms of context being used in most of existing methods. To address this, we formulate a flexible, discriminative model that is able to learn both bias and latent factors by considering such features, resulting in a better understanding of movie preferences and improved recommendation performance.

The second part of this dissertation tackles the problem of corporating cross-domain context in recommendation. In this direction, transfer learning techniques have demonstrated promising successes. However, despite much encouraging progress, most of the advances in transfer learning still take place in the condition of fully entity correspondences between two domains. The nature of cross-domain recommendation compels us to move beyond the existing paradigm of transfer learning to develop novel and more generalized algorithms. Towards this end, we build methods and techniques for general transfer learning in cross-domain recommendation settings, which enable to construct entity correspondence with limited budget by using active learning strategy to facilitate knowledge transfer across domains. In particular, first we propose a unified framework for cross-domain recommendation. It allows us to identify correspondences that can bring as much knowledge as possible, then we can conduct efficient transfer model to improve recommendation quality. Second, based on the framework, we develop three active learning solutions that can iteratively select entities in target domain to query their correspondences from the source. After that, we embed the correspondences into three novel transfer learning models respectively. We demonstrate that these solutions can take advantage of both active learning and transfer learning techniques, lead to many practical benefits.