Harmonized recommendation with large language model guidance and reinforcement learning

Abstract

Recommender systems have become critical infrastructure in engineering applications such as e-commerce, content delivery platforms, and digital service environments. Despite their effectiveness, existing graph neural network (GNN) based recommenders face three unresolved technical gaps: (i) semantic poverty, where sparse identifier embeddings provide insufficient signal for cold-start and long-tail items; (ii) contrastive learning instability, where false negative sampling and fixed temperature parameters fail to adapt to heterogeneous user–item degree distributions; and (iii) underutilized large language model (LLM) reasoning, where existing methods treat large language models (LLMs) as static feature extractors, neglecting their chain-of-thought reasoning capacity for structured semantic understanding. To address these challenges, we propose HarmoRec, a harmonized recommendation framework integrating three tightly coupled components. First, the Deep Semantic Embedding Generator (DSEG) leverages fine-tuned large language models (LLMs) to produce chain-of-thought item rationales encoded as dense semantic identifiers without introducing online latency. Second, the Reinforcement-Driven Contrastive Learning (RDCL) module reformulates negative sampling and temperature tuning as an adaptive reinforcement learning (RL) task. Third, the Harmonized Group Policy Optimization (HGPO) mechanism coordinates learning across degree-based node groups to improve long-tail recommendation equity. Experiments on four real-world engineering datasets for recommender systems demonstrate improvements of 8%–10% over strong baselines, with gains of 17%–20% on long-tail items (p<0.001, Cohen’s d≥9.85), while maintaining real-time inference efficiency (28–30 ms per batch). These results highlight the effectiveness of combining large language models and reinforcement learning for practical recommender system applications.