Bridging the User-side Knowledge Gap in Knowledge-aware Recommendations with Large Language Models (2412.13544v1)

Abstract: In recent years, knowledge graphs have been integrated into recommender systems as item-side auxiliary information, enhancing recommendation accuracy. However, constructing and integrating structural user-side knowledge remains a significant challenge due to the improper granularity and inherent scarcity of user-side features. Recent advancements in LLMs offer the potential to bridge this gap by leveraging their human behavior understanding and extensive real-world knowledge. Nevertheless, integrating LLM-generated information into recommender systems presents challenges, including the risk of noisy information and the need for additional knowledge transfer. In this paper, we propose an LLM-based user-side knowledge inference method alongside a carefully designed recommendation framework to address these challenges. Our approach employs LLMs to infer user interests based on historical behaviors, integrating this user-side information with item-side and collaborative data to construct a hybrid structure: the Collaborative Interest Knowledge Graph (CIKG). Furthermore, we propose a CIKG-based recommendation framework that includes a user interest reconstruction module and a cross-domain contrastive learning module to mitigate potential noise and facilitate knowledge transfer. We conduct extensive experiments on three real-world datasets to validate the effectiveness of our method. Our approach achieves state-of-the-art performance compared to competitive baselines, particularly for users with sparse interactions.

Bridging the User-side Knowledge Gap in Knowledge-aware Recommendations with LLMs

The paper addresses the challenge of leveraging user-side structural knowledge in recommender systems, a critical aspect often overlooked in the field of knowledge-aware recommendation. Traditional systems have successfully integrated item-side knowledge derived from Knowledge Graphs (KGs) to enhance recommendation accuracy. However, obtaining comparable insights from user-side data remains challenging due to the scarcity and improper granularity of existing user-side features. This research proposes a novel method to bridge this gap using LLMs.

Key Contributions

1. LLM-based User-side Knowledge Inference: The research introduces an inferential method using LLMs to generate structured knowledge of user interests from past behaviors. By deploying LLMs, the method captures abstract and nuanced user interests, which are inherently difficult to model with conventional approaches. This knowledge is organized into a structured format and integrated with existing item-side data, thus allowing the construction of a Comprehensive Interest Knowledge Graph (CIKG).
2. CIKG-based Recommendation Framework: To exploit the inferred user-side knowledge and enhance recommendation accuracy, the paper presents a novel recommendation framework. This framework employs Graph Neural Networks (GNNs) to process the hybrid structure formed by the CIKG. Two major innovations in this framework are:
   • A User Interest Reconstruction Module based on a Graph Masked Autoencoder (GMAE), designed to enhance the robustness of the model against noise potentially introduced by LLM hallucinations.
   • A Cross-domain Contrastive Learning Module to facilitate effective transition of knowledge from the auxiliary information domain to the recommendation domain, ensuring seamless integration of user-side and item-side knowledge.
3. Empirical Evaluation: Extensive experiments across three real-world datasets demonstrate the framework's efficacy, achieving state-of-the-art performance compared to existing methods. The approach is particularly effective for users with sparse interaction histories, showcasing its capability in addressing data sparsity issues.

Technical Insights and Results

The core technical innovation lies in the transformation of natural language inference into a form that harmonizes with graph-based recommendation. The method abstracts user interests into structured nodes that offer more precise user modeling, an approach validated against competitive baselines. The incorporation of both user-side and item-side data in the unified CIKG structure signifies a leap in the ability to mine relationships and patterns traditionally difficult to access.

The paper presents notable numerical advancements in recommendation metrics such as Recall and NDCG, particularly for users in sparse interaction scenarios. These results underscore the value of incorporating LLMs in expanding the horizon of user modeling, effectively tackling inherent limitations posed by cold starts and data sparsity.

Implications and Future Work

This research extends the application of LLMs beyond their conventional domains, illustrating their potential in enhancing the fidelity of user interest representations in recommendation systems. From a theoretical standpoint, the integration of LLMs in the construction of user-side structural knowledge paves the way for more holistic and detailed preference modeling.

Practically, this enhancement can lead to more personalized and accurate recommendations which are crucial in domains such as e-commerce, content streaming, and social media. Future explorations could include refining LLM inference to further minimize noise and exploring the scalability of such systems across larger and more diverse datasets. Additionally, further work could investigate alternative methods of structuring LLM output into graph-compatible formats, enabling broader applications in knowledge-based AI architectures.