Deep Learning on Recommender Systems

The so-called cold-start problem has haunted the recommender systems community for years. The problem happens when a user rated/clicked/liked a few number of items. Classic approaches, such as collaborative filtering, assume that a user has a fair amount of actions so that the preferences of the user can be inferred. As a result, traditional methods cannot effectively model the interests of cold users due to the scarcity of data. In this dissertation, I will introduce our recent works on employing deep learning methods for alleviating the cold-start problem for recommendation. In the first part, we focus on utilizing review data to build deep learning models to ease the cold start problem. In the second part, I present an spectral approach to discover users' interests from the spectral domain of the user-item bipartite graph. In the third part, I introduce a recurrent method designed to capture users' evolving interests from dynamic graphs. In the fourth part, we propose to model users and items with probability distributions, rather than the popular vectors. With distribution-based representations, the proposed model is able to alleviate the cold-start problem and therefore, delivers the start-of-the-art performances in three real-world datasets.