Review-Enhanced Cross-Domain Recommendation using Graph Representations and Large Language Models
MSc Thesis Defense by: Sepideh Ahmadian
Date: 5th of February
Time: 1:00 PM
Location: EH 122
Abstract:
A major challenge in recommender systems is data sparsity. Cross-domain recommendation (CDR) frameworks address this by transferring knowledge from high-resource to low-resource domains, but existing methods largely rely on user ratings that provide only implicit preference signals and may overlook subtle user preferences. In this work, we propose a novel review- based CDR framework that leverages Large Language Models (LLMs) to extract fine-grained product aspects and associateduser sentiments from reviews. This approach enables the capture of nuanced and explicit user preferences that go beyond traditional rating-based methods. We construct comprehensive user profiles by aggregating aspect-level features across both source and target domains. The relationships among users, items, and extracted aspects are then jointly modeled using graph-based representations. We explore multiple graph architectures to identify the most effective approach. The first is a knowledge graph with users, features, and products as nodes. The second is a weighted bipartite user-product graph, wherefeatures are encoded as edge attributes. The third is a hypergraph representation, where each hyperedge explicitly connects a user, an item, and the corresponding extracted aspects, enabling a unified representation of their interdependent relationships. These representations are used to train our recommender system, implemented with graph attention networksand hypergraph neural networks. Extensive experiments demonstrate the effectiveness of our approach for personalized recommendations in data-sparse settings. The knowledge graph and hypergraph formulations consistently outperform strong baseline methods across multiple recommendation metrics, while the weighted graph variant achieves comparable performance with lower computational cost. Overall, our framework combines traditional rating-based CDR with semantic understanding of user-generated content through LLM-extracted aspects, offering a robust solution for effective preference transfer and personalized recommendations through shared semantic representations.
Thesis Committee:
Internal Reader: Dr. Alioune Ngom
External Reader: Dr. Huapeng Wu
Advisor: Dr. Luis Rueda
Chair: Dr. Muhammad Asaduzzaman