Download files
Abstract
Recommender system (RS) is a useful information filtering tool for guiding users in a personalized way of discovering products/services from a large space of possible options. A good recommender system can not only ameliorate the prevalent issue of over-choice due to information explosion but can also promote sales and boost revenues. As such, it has become a vital and indispensable component in the modern internet industry.
In this thesis, we aim to tackle several key prevalent tasks in recommender systems, including (1) modeling the complex relationships between users and items; (2) integrating side information; (3) modeling the temporal dynamics; and (4) cold-start problem. To this end, we investigate neural architectures and geometric methods, aiming to tackle those challenges and improve recommendation performances.
Recently, deep learning's revolutionary advances in speech recognition, computer vision, and natural language processing have gained significant attention. Many challenging problems such as feature representation learning and sequence modeling can be well addressed with deep neural models. Our key contributions on personalized recommendations with neural architectures are listed as follows: Firstly, we investigated feature representation learning for hybrid recommender systems, which involves learning items' feature representation with a contractive autoencoder. We proposed AutoSVD and AutoSVD++ to efficiently anticipate user's ratings for unseen items. Secondly, we investigated the personalized ranking task and proposed NeuRec and its extension, SNR, to model the complex and nonlinear relationships between users and items. Thirdly, we investigate the challenging cold-start problem and present a model that considers both item and user side information under a unified deep neural framework which can introduce personalization and collaborative effects to cold-start recommendations. Fourthly, we proposed a self-attentive metric learning framework to tackle the temporal dynamics in recommender systems.
Beyond neural networks, we also investigate geometric methods which concern modeling the intricate relations between users and items with geometric inductive biases. We proposed FML to factorize the explicit Euclidean distance with standard learning algorithms. We also proposed a collaborative filtering model using Quaternion representations. Those methods achieve comparable performances as neural networks with simpler formulation.