New Social Collaborative Filtering Algorithms for Recommendation on Facebook
Joseph R. Christian, G. Noel · 2011
This thesis examines the problem of designing efficient, scalable, and accurate social collaborative filtering (CF) algorithms for personalized link recommendation on Facebook. Unlike standard CF algorithms using relatively simple user and item features (possibly just the user ID and link ID), link recommendation on social networks like Facebook poses the more complex problem of learning user preferences from a rich and complex set of user profile and interaction information. Most existing social CF (SCF) methods have extended traditional CF matrix factorization (MF) approaches, but have overlooked important aspects specific to the social setting; specifically, existing SCF MF methods (a) do not permit the use of item or link features in learning user similarity based on observed interactions, (b) do not permit directly modeling useruser information diffusion according to the social graph structure, and (c) cannot learn that that two users may only have overlapping interests in specific areas. This thesis proposes a unified SCF optimization framework that addresses (a)–(c) and compares these novel algorithms with a variety of existing baslines. Evaluation is carried out via live user trials in a custom-developed Facebook App involving data collected over three months from over 100 App users and their nearly 30,000 friends. Not only do we show that our novel proposals to address (a)–(c) outperform existing approaches, but we also identify which offline ranking and classigication evaluation metrics correlate most with human judgment of algorithm performance. Overall, this thesis represents a critical step forward in extending SCF recommendation algorithms to fully exploit the rich content and structure of social networks like Facebook.