Weak Signals from the Web: New Perspectives in Collaborative Filtering

EDIC candidacy exam
Exam president: Prof. Robert West
Thesis advisor: Prof. Karl Aberer
Co-examiner: Prof. Pierre Dillenbourg

Abstract
Personalized recommendation has become a crucial aspect of many information systems, as it increases content visibility, creates users engagement and facilitates navigation. With their growing presence and the broadening of their applications, recommender systems are exposed to an ever-growing amount of users. 
Traditional methods, relying on user ratings as a primary source of information, are unable to produce recommendations for people that have never evaluated any products. 
Based on this observation, new methods have been proposed to provide accurate results from observations coming from users' natural activity. In this work, we discuss three examples of recommender systems that rely on users implicit feedbacks and take advantage of external sources of information. We first discuss an optimization procedure to infer a preference structure from positive-only user interactions. We then discuss a method to rely on the social graph to circumvent high data sparsity. Last, we examine a method to alleviate the cold-start problem by using visual features.

Background papers
BPR: Bayesian personalized ranking from implicit feedback, Rendle et al.
[UAI 09] VBPR: Visual Bayesian Personalized Ranking from Implicit Feedback, He et al.
[AAAI 16] Recommender Systems with Social Regularization, Ma et al. [WSDM 11]