Improving Collaborative Video Recommendation: Methods Addressing the Data Sparsit

Recommender systems are becoming increasingly indispensable nowadays in solving the information overload problem and providing users with personalized information and services, such as video recommendation in Youtube and Netflix. Collaborative Filtering (CF) is one of the most widely used techniques, which predicts a user's interest from feedbacks of similar users or items. However, it is challenging to achieve accurate recommendation for users with little or no historical feedback, known as data sparsity problem. The aim of this thesis is to improve the performance of collaborative (CF-based) video recommendation by addressing the data sparsity problem.;In this thesis, we first propose a social-group-based algorithm to produce personalized video recommendation by ranking candidate videos from the groups a user is affiliated with. We implement the algorithm in the Tencent Video service system, and the online A/B testing results show that the proposed algorithm not only improves the click-through rate, but also recommends more diverse videos.;Then, we propose another method to enhance the neighborhood-based collaborative filtering by using a hybrid tree-encoded linear model to predict how like-minded two users are in viewing videos, based on their demographic, social, and temporal information. Besides, tag-based user pro ling is adopted to measure user-user similarity. The experimental results validate the strength of our prediction model and tag-based user pro ling scheme in practical recommender systems.;Thirdly, we propose a transfer learning model by using multisite information to address the data sparsity problem. Specifically, we propose a generative model of Multi-site Probabilistic Factorization (MPF) to model user preferences in multiple video websites. Extensive recommendation validation shows that the MPF model can significantly improve recommendation accuracy compared to several other state-of-the-art factorization models. Our findings provide insights on the value of integrating user data from multiple sites, which stimulates potential collaborations between video service providers. Furthermore, to justify the general applicability of this model and understand the underlying market forces that lead to what we observe, we model user demand and ISP strategies based on evidence from data, and arrive at a two-CP-two-video-type game that can be used to explain the observed market equilibrium with multiple service providers.