CHIME: A Compressive Framework for Holistic Interest Modeling

2025-04-09

AI Summary

To address the high computational cost and difficulty of modeling full behavioral contexts in recommender systems, this paper proposes the Compressed Integrated Interest Modeling (CIM) framework. CIM is the first to jointly integrate multi-granularity contrastive learning—capturing both persistent and transient user interests—with residual vector quantization (RVQ) to produce compact, high-fidelity interest embeddings. It further leverages large language models to encode long, heterogeneous behavioral sequences and fuses multi-source information. The design ensures real-time inference capability and system scalability while preserving critical interest signals. Extensive experiments on multiple public benchmarks demonstrate consistent and significant improvements in ranking metrics—including NDCG@10 and Recall@20—validating CIM’s efficiency, effectiveness, and strong generalization across diverse domains.

Technology Category

• Machine Learning: Learning Preferences or Rankings
• Data Mining & Knowledge Management: Recommender Systems
• Cognitive Modeling & Cognitive Systems: Simulating Human Behavior

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs

Abstract

Modeling holistic user interests is important for improving recommendation systems but is challenged by high computational cost and difficulty in handling diverse information with full behavior context. Existing search-based methods might lose critical signals during behavior selection. To overcome these limitations, we propose CHIME: A Compressive Framework for Holistic Interest Modeling. It uses adapted large language models to encode complete user behaviors with heterogeneous inputs. We introduce multi-granular contrastive learning objectives to capture both persistent and transient interest patterns and apply residual vector quantization to generate compact embeddings. CHIME demonstrates superior ranking performance across diverse datasets, establishing a robust solution for scalable holistic interest modeling in recommendation systems.

Problem

• Modeling holistic user interests with high computational cost
• Handling diverse information with full behavior context
• Overcoming signal loss in existing search-based methods

Innovation

• Adapted large language models for behavior encoding
• Multi-granular contrastive learning for interest patterns
• Residual vector quantization for compact embeddings