Toward Intelligent Personalization: Integrating Large Language Models with Recommender Systems
PhD Dissertation Defense by: Danial Ebrat
Date: May 19th, Tuesday
Time: 12:00 PM
Teams Link: https://teams.microsoft.com/meet/286038174789864?p=C63FbrDYh1YKLM8ES0
Abstract:
Recommender systems power the digital services we use daily, yet they struggle with sparse data, new users, and opaque decisions. This thesis investigates how Large Language Models (LLMs) can be systematically integrated with graph-based collaborative filtering to build recommender systems that are simultaneously more accurate, robust, and interpretable. The proposed framework transforms rich textual metadata into semantic embeddings that initialize user–item graphs, propagated through Graph Attention Networks to blend meaning with structure. A modular agent-based architecture coordinates semantic profiling, ranking, reranking via a confidence-weighted search tree, and natural-language explanation generation. A custom hybrid loss aligns semantic and interaction-based signals, with particular gains for users who have little interaction history. The thesis also introduces Lusifer, a simulation environment that generates realistic evolving user feedback for controlled training and evaluation. Experiments across movie, book, and music datasets show consistent improvements over state-of-the-art baselines, especially under cold-start and data-sparse conditions.
Keywords:
Recommender Systems, Large Language Models, Vector Embeddings
Doctoral Committee:
Internal Reader: Dr. Jianguo Lu
Internal Reader: Dr. Ikjot Saini
External Reader: Dr. Esraa Abdelhalim
External Examiner: Dr. Miguel Vargas Martin
Advisor(s): Dr. Luis Rueda
Chair: Dr. Beth-Anne Schuelke-Leech