Synergistic Integration and Discrepancy Resolution of Contextualized Knowledge for Personalized Recommendation

TL;DR

The paper addresses the challenge of effectively integrating large language model (LLM) world knowledge into personalized recommender systems (RSs). It introduces CoCo, an end-to-end collaboration-contradiction fusion framework that dynamically generates user-specific semantic knowledge via a vector-quantized soft-prompt pool and aligns semantic signals with behavioral representations through cross-attention, complemented by a contradiction-elimination stage that uses LoRA fine-tuning and gradient masking. Key contributions include (i) a novel prompt collaboration mechanism with a discrete codebook and per-user prompt selection, (ii) semantic decoupling via an orthogonal constraint to reduce prompt redundancy, (iii) semantic-behavioral alignment to fuse signals even across diverse backbones, and (iv) a dynamic contradiction-elimination strategy validated on public and industrial datasets, showing up to 8.58% improvements in ranking metrics and online revenue gains of 1.91% in production. CoCo demonstrates robust applicability across RS backbones and scales with larger LLMs, offering practical pathways for deployed, knowledge-enhanced, personalized recommendations.

Abstract

The integration of large language models (LLMs) into recommendation systems has revealed promising potential through their capacity to extract world knowledge for enhanced reasoning capabilities. However, current methodologies that adopt static schema-based prompting mechanisms encounter significant limitations: (1) they employ universal template structures that neglect the multi-faceted nature of user preference diversity; (2) they implement superficial alignment between semantic knowledge representations and behavioral feature spaces without achieving comprehensive latent space integration. To address these challenges, we introduce CoCo, an end-to-end framework that dynamically constructs user-specific contextual knowledge embeddings through a dual-mechanism approach. Our method realizes profound integration of semantic and behavioral latent dimensions via adaptive knowledge fusion and contradiction resolution modules. Experimental evaluations across diverse benchmark datasets and an enterprise-level e-commerce platform demonstrate CoCo's superiority, achieving a maximum 8.58% improvement over seven cutting-edge methods in recommendation accuracy. The framework's deployment on a production advertising system resulted in a 1.91% sales growth, validating its practical effectiveness. With its modular design and model-agnostic architecture, CoCo provides a versatile solution for next-generation recommendation systems requiring both knowledge-enhanced reasoning and personalized adaptation.

Figures (9)

• Figure 1: The problems of the existing paradigm and the challenges of our framework.
• Figure 2: The performance comparison across 4 distinct prompts: Prompt 1 denotes the universal prompt, Prompt 2 is guided by age-based features, Prompt 3 is guided by gender-based features, and Prompt 4 is guided by item category.
• Figure 3: (a) Performance gains of RSs from LLM-generated knowledge across 5 user groups (left). (b) Visualization of the semantic space represented by LLMs and the behavioral space represented by RSs (right).
• Figure 4: The overview of CoCo: An end-to-end collaboration-contradiction fusion framework integrating large language models and recommendation systems. In the collaboration enhancement phase, we dynamically generate semantic knowledge for each user that is highly aligned with their behavioral patterns. In the contradiction elimination phase, we achieve alignment between the behavioral space and the semantic space through user-level LoRA fine-tuning.
• Figure 5: Sensitivity Analysis of the prompt candidate size $K$.