Improving Content-Aware Recommendation: Relationship Mining, Integrating, and Disentangling
Abstract
Explosive growth in the amount of information has infiltrated every aspect of our lives. Recommender systems, as an effective tool to filter information, have become prevalent and influential in many domains. This dissertation seeks to improve recommender systems by learning the unique properties and complex relationships of the rich content information associated with both users (who seek recommendations) and items (that are recommended to users). However, to do so, there are key research challenges: (i) the extreme sparsity of the observed item relations; (ii) high het-erogeneity between user-user relations and user-item interactions; (iii) duplication between content and collaborative signals; and (iv) the highly skewed long-tail distribution of user feedback towards items. With these challenges, this dissertation research makes four unique contributions:
• The first research contribution of this dissertation is to mine and integrate the item-side relations. Concretely, we begin by investigating different types of item relations (e.g., complementary and substitute relations), and build a novel neural item-relationship based model to uncover the item relationships. We then integrate these item multi-relations into user sequences through a hierarchical temporal graph to enhance sequential recommendation.
• The second research contribution of this dissertation is to mine and integrate the user-side relations. We are one of the first to investigate the socio-behavioral phenomenon of social resonance to closely connect user relations with user interactions towards items. Then we integrate the user social relations in sequential recommendations to improve fashion recommendations. We also provide a new large visual dataset of dynamic visual posts by key fashion bloggers that contain both high-quality dynamic fashion preference shifts over time.
• The third research contribution of this dissertation is to disentangle the content and collab-orative features to address the duplication problem in content-aware recommendations. We propose a novel two-level disentanglement approach that supports both content-collaborative disentanglement and feature disentanglement based on a variational auto-encoder.
• The fourth research contribution of this dissertation is to decouple the learning process that tackles the skewed long-tail distribution problem. We propose a novel cross decoupling framework that utilizes cumulative learning and multi-experts to decouple the learning of long-tail distribution from prior and conditional knowledge aspects.
Subject
Recommender SystemsMachine Learning
Disentangling
Relationship Mining
Fashion Recommendation
Citation
Zhang, Yin (2022). Improving Content-Aware Recommendation: Relationship Mining, Integrating, and Disentangling. Doctoral dissertation, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /197113.