Research Area:  Machine Learning

Abstract:

  The widespread adoption of the Internet has led to an explosion in the number of choices available to consumers. Users begin to expect personalized content in modern E-commerce, entertainment and social media platforms. Recommender Systems (RS) provide a critical solution to this problem by maintaining user engagement and satisfaction with personalized content.
  Traditional RS techniques are often linear limiting the expressivity required to model complex user-item interactions and require extensive handcrafted features from domain experts. Deep learning demonstrated significant breakthroughs in solving problems that have alluded the artificial intelligence community for many years advancing state-of-the-art results in domains such as computer vision and natural language processing. The recommender domain consists of heterogeneous and semantically rich data such as unstructured text (e.g. product descriptions), categorical attributes (e.g. genre of a movie), and user-item feedback (e.g. purchases). Deep learning can automatically capture the intricate structure of user preferences by encoding learned feature representations from high dimensional data. In this thesis, we explore five novel applications of deep learning-based techniques to address top-n recommendation. First, we propose Collaborative Memory Network, which unifies the strengths of the latent factor model and neighborhood-based methods inspired by Memory Networks to address collaborative filtering with implicit feedback. Second, we propose Neural Semantic Personalized Ranking, a novel probabilistic generative modeling approach to integrate deep neural network with pairwise ranking for the item cold-start problem. Third, we propose Attentive Contextual Denoising Autoencoder augmented with a context-driven attention mechanism to integrate arbitrary user and item attributes. Fourth, we propose a flexible encoder-decoder architecture called Neural Citation Network, embodying a powerful max time delay neural network encoder augmented with an attention mechanism and author networks to address context-aware citation recommendation. Finally, we propose a generic framework to perform conversational movie recommendations which leverages transfer learning to infer user preferences from natural language. Comprehensive experiments validate the effectiveness of all five proposed models against competitive baseline methods and demonstrate the successful adaptation of deep learning-based techniques to the recommendation domain.

Name of the Researcher:  Travis Akira Ebesu

Name of the Supervisor(s):  Yi Fang

Year of Completion:  2019

University:  Santa Clara University

Thesis Link:   Home Page Url