Review-Enhanced Cross-Domain Recommendation using Graph Representations and Large Language Models
MSc Thesis Proposal by: Sepideh Ahmadian
Date: Wednesday, November 12, 2025
Time: 1:00 pm
Location: Essex Hall, Room 122
A major challenge in recommender systems (RS) is cold start and data sparsity. Cross-domain recommendation (CDR) frameworks address these issues by transferring knowledge from high-resource (HR) to low-resource (LR) domains, often using user ratings. However, these ratings 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 associated user sentiments from reviews. This approach enables the capture of nuanced and explicit user preferences. We construct comprehensive user profiles by aggregating features across both source and target domains. The relationships between users, features, and products are then modelled using graph-based representations. We explore two graph architectures: the first is a knowledge graph with users, features, and products as nodes; the second is a weighted bipartite user–product graph, where features are encoded as edge attributes. These representations are used to train our recommender system, which is implemented with a graph attention network. Extensive experiments demonstrate the effectiveness of our approach for personalized recommendations. The knowledge graph consistently outperforms baseline methods across multiple recommendation metrics, while the weighted graph variant achieves comparable performance to the knowledge graph with lower computational cost. Overall, our framework combines traditional rating-based CDR with the semantic understanding of user-generated content, offering a robust solution for personalized recommendations in data-sparse settings.
Internal Reader: Dr. Alioune Ngom
External Reader: Dr. Huapeng Wu
Advisor(s): Dr. Dima Alhadidi and Dr. Luis Rueda
