A thesis submitted in partial fulfillment of the requirements for the degree of Master of Science in Data Science and Artificial Intelligence

Thesis (M. Sc.) - Asian Institute of Technology, 2025

Recommender systems (RS) are essential for delivering personalized recommendations, enhancing user experiences, and increasing business profitability. Traditional RS methods, such as collaborative filtering, content-based filtering, and hybrid methods, offer unique benefits but also face certain constraints. Collaborative filtering, exemplified by LightGCN, excels at capturing topological patterns from user-item interactions but struggles with limited user-item interactions and cold-start problems. Content-based filtering mitigates item cold-start issues by leveraging item features but overlooks collaborative preferences from similar users. To address these challenges, this study proposes GSAGE-LGCN, a hybrid model integrating GraphSAGE with LightGCN. GraphSAGE enhances node embeddings through inductive feature learning from sampled neighbors, complementing LightGCN{u2019}s collaborative filtering capabilities. Three variants (V1, V2, V3) were developed to explore different integration strategies, evaluated on MovieLens-100K, MovieLens-1Mil, Brazilian E-Commerce, and Amazon Book 2018 datasets. Results show that GSAGE-LGCN V1 significantly improves performance in low-density and cold-start scenarios, achieving a 62.1% and 97.6% increase in Precision@20 on MovieLens 100K and MovieLens-1Mil datasets respectively under user cold-start conditions. In the very low density dataset, Brazilian E-Commerce, V1 achieves a 96.3% increase in Precision@20 in the cold-start scenario. This research highlights the potential of combining GraphSAGE and LightGCN to address the issues arising from limited user-item interactions and cold-start prob lems, offering a robust hybrid recommender system for diverse recommendation scenarios.