Xiang Wang

[advanced] Large Language Models for User Behavior Modeling: Cross-Modal Interpretation, Preference Optimization, and Agentic Simulation

Summary

This lecture examines the pivotal role of large language models (LLMs) in advancing user behavior modeling, with a specific focus on personalized recommendation systems. It highlights how LLMs enhance the understanding of user intent, optimize behavioral preferences, and simulate intricate user interactions, with three foundational pillars:

1. Cross-Modal Interpretation: This section explores how LLMs process diverse behavioral modalities (e.g., clicks, purchases, and views) alongside textual data, enabling a unified understanding of user behaviors across multiple input formats.
2. Preference Alignment: It introduces state-of-the-art approaches such as reinforcement learning with human feedback (RLHF) and direct preference optimization (DPO), framing user behaviors as preference data to refine and personalize recommendations effectively.
3. Agentic Simulation: The lecture further discusses decomposing complex user modeling tasks into manageable subtasks, each addressed by specialized LLMs operating collaboratively within a network of agentic experts.

The lecture provides both theoretical insights and practical applications, bridging the gap between LLMs research and real-world recommender systems.

Syllabus

• Background: role of LLMs in recommendation systems.
• Introduction: key pillars (cross-modal interpretation, preference optimization, agentic simulation).
• Cross-Modal Interpretation: concepts, paradigms, and core ideas of fine-tuning.
• Preference Alignment: concepts, methodologies, and recent progress of preference optimization.
• Agentic Simulation: concepts, frameworks, and latest advancements of agents.
• Key takeaways and future trends

References

LLaRA: Large Language-Recommendation Assistant. In SIGIR 2024.

On generative agents in recommendation. In SIGIR 2024.

On Softmax Direct Preference Optimization for Recommendation. In NeurIPS 2024.

Customizing Language Models with Instance-wise LoRA for Sequential Recommendation. In NeurIPS 2024.

beta-DPO: Direct Preference Optimization with Dynamic beta. In NeurIPS 2024.

Language Representations Can Be What Recommenders Need: Findings and Potentials.

Pre-requisites

Machine learning basics.

Short bio

Xiang Wang is a professor at the University of Science and Technology of China (USTC). He obtained his Ph.D. from the National University of Singapore in 2019. His research focuses on multi-modal large language models, user modeling, and data mining. He has authored over 100 publications in top-tier artificial intelligence conferences, including more than 10 papers recognized as highly influential and 3 papers selected as Best Paper Finalists. He has served as an Area Chair for leading conferences such as NeurIPS and ICML. Dr. Wang’s achievements include the SIGIR Early Career Award and recognition as one of MIT Technology Review’s Innovators Under 35 (TR35) in China.