A Unified Approach to Collaborative Filtering via Linear Models and Beyond

Abstract

Recommending a personalised list of items to users is a core task for many online services such as Amazon, Netflix, and Youtube. Recommender systems are the algorithms that facilitate such personalised recommendation. Collaborative filtering (CF), the most popular class of recommendation algorithm, exploits the wisdom of the crowd by predicting users’ preferences not only from her past actions but also from the preferences of other like-minded users. In general, it is desirable to have a CF framework that is (1) applicable to wide range of recommendation scenarios, (2) learning-based, (3) amenable to convex optimisation, and (4) scalable. However, all existing CF methods, such as neighbourhood and matrix factorisation, lack one or more of these desiderata.

In this dissertation, we investigate linear models, an under-appreciated but promising area for recommendations that addresses all the above desiderata. We formulate a unified framework based on linear models that yields CF algorithms for four prevalent scenarios. First, we investigate Social CF, which involves leveraging users’ signals from online social networks. We propose social affinity filtering (SAF), that exploits fine-grained user interactions and activities in a social network. Second, we investigate Cold-Start CF, which refers to the scenario when we do not have any historical data about a user or an item. We formulate a large-scale linear model that leverages users social information. Third, we investigate One-Class CF, which concerns suggesting relevant items to users from the data that consists of only positive preferences such as item purchase.Noting the superior performance of linear models, we propose LRec, a user-focused linear CF model, and extend it to large-scale datasets via dimensionality reduction. Finally, we investigate Explicit Feedback CF, which concerns predicting user’s actual preferences such as rating, or like/dislikes. We identify CF as an auto-encoding problem and propose AUTOREC, a generalized neural network architecture for CF. We demonstrate state-of-the-art performance of the proposed models through extensive experimentation on real world datasets.

In a nutshell, this dissertation elucidates the power of linear models for various CF tasks and paves the way for further research on applying deep learning models to CF.