Author

Wei Mo

Date of Award

2025

Document Type

Dissertation - SCU Access Only

Publisher

Santa Clara : Santa Clara University, 2025

Degree Name

Master of Science (MS)

Department

Computer Science and Engineering

First Advisor

Yi Fang

Abstract

Most research in recommender systems focuses on relevance, but diversity is equally important as it helps prevent filter bubbles and provides users with meaningful choices. This thesis explores the application of LoRA-enhanced LLMs for recommendation tasks across multiple datasets. While Large Language Models (LLMs) offer strong reasoning capabilities, standard LoRA-based fine-tuning often prioritizes personalization and relevance, struggling to maintain diversity and global quality in recommendations.

To address this, we propose two key enhancements:

(1) Personalized Adaptive Negative Sampling (PANS), dynamically balancing same-genre vs. different-genre exploration based on user engagement.

(2) Multi-Objective Loss Optimization, incorporating User Preference Score, Global Quality Score, and Genre-Based Entropy Loss to generate diverse and high-quality recommendations.

We further employ Bayesian Optimization for efficient hyperparameter tuning, ensuring faster convergence and better trade-offs between personalization, quality, and diversity. Extensive experiments on multiple datasets demonstrate the effectiveness of our approach in achieving a balance between recommendation accuracy and diversity.

Share

COinS