LLM-CoT Enhanced Graph Neural Recommendation with Harmonized Group Policy Optimization

TL;DR

This paper addresses information sparsity in graph-based recommenders by leveraging offline LLM chain-of-thought (CoT) reasoning to produce semantic IDs that enrich item representations. It introduces LGHRec, which fuses CoT-derived semantic IDs with ID embeddings to initialize GNN representations and employs Harmonized Group Policy Optimization (HGPO) to adaptively optimize graph contrastive learning via negative sampling and temperature across degree-based groups. The approach yields consistent improvements across Yelp2018, Amazon-Book, and MIND, particularly boosting long-tail recommendations and stabilizing cross-group performance, while maintaining efficient online inference due to offline LLM preprocessing. The findings highlight the practical potential of integrating LLM-CoT reasoning with reinforcement-learning-guided contrastive learning for scalable, information-dense recommender systems and point to multimodal extensions as promising future work.

Abstract

Graph neural networks (GNNs) have advanced recommender systems by modeling interaction relationships. However, existing graph-based recommenders rely on sparse ID features and do not fully exploit textual information, resulting in low information density within representations. Furthermore, graph contrastive learning faces challenges. Random negative sampling can introduce false negative samples, while fixed temperature coefficients cannot adapt to the heterogeneity of different nodes. In addition, current efforts to enhance recommendations with large language models (LLMs) have not fully utilized their Chain-of-Thought (CoT) reasoning capabilities to guide representation learning. To address these limitations, we introduces LGHRec (LLM-CoT Enhanced Graph Neural Recommendation with Harmonized Group Policy Optimization). This framework leverages the CoT reasoning ability of LLMs to generate semantic IDs, enriching reasoning processes and improving information density and semantic quality of representations. Moreover, we design a reinforcement learning algorithm, Harmonized Group Policy Optimization (HGPO), to optimize negative sampling strategies and temperature coefficients in contrastive learning. This approach enhances long-tail recommendation performance and ensures optimization consistency across different groups. Experimental results on three datasets demonstrate that LGHRec improves representation quality through semantic IDs generated by LLM's CoT reasoning and effectively boosts contrastive learning with HGPO. Our method outperforms several baseline models. The code is available at: https://anonymous.4open.science/r/LLM-Rec.

Figures (8)

• Figure 1: The differences between the three recommendation paradigms.
• Figure 2: The architecture diagram of the proposed LGHRec.
• Figure 3: After using various fine-tuning methods, the general capabilities and recommendation performance of LGHRec: (a)Yelp dataset, (b)MIND dataset.
• Figure 4: The prompt template for guiding LLM to perform CoT reasoning.
• Figure 5: A comparison of NDCG@20 between LGHRec and baseline models across three datasets, grouped by different user and item interaction levels based on interaction count.