Title: Maintaining Collaborative Frequency Components in LLM-Based Recommendations: A Step Beyond Semantic Comprehension

Abstract:

The integration of Large Language Models (LLMs) with recommender systems offers a promising pathway for generating recommendations grounded in semantic understanding. Nevertheless, LLM-driven recommenders often exhibit a bias toward amplifying semantic correlations found in user interaction histories. Unlike conventional Transformer-based sequential models, which typically preserve or even strengthen collaborative signals to achieve state-of-the-art results, LLM-based architectures tend to dilute the intrinsic collaborative cues embedded in pretrained collaborative ID embeddings as they traverse the model’s layers.

To overcome this challenge, we propose FreLLM4Rec, a novel framework that harmonizes semantic and collaborative data through a spectral lens. Initially, item embeddings—rich with both semantic and collaborative attributes—are refined via a Global Graph Low-Pass Filter (G-LPF) to strip away extraneous high-frequency noise. Subsequently, Temporal Frequency Modulation (TFM) actively safeguards collaborative signals at each layer. Theoretically, the ability of TFM to preserve these signals is validated by linking the optimal but computationally prohibitive local graph Fourier filters to suboptimal yet efficient frequency-domain filters.

Our extensive evaluation across four benchmark datasets confirms that FreLLM4Rec effectively counters the attenuation of collaborative signals, delivering competitive outcomes. Notably, it yields performance gains of up to 8.00% in NDCG@10 compared to the strongest baseline. These results not only shed light on the mechanisms by which LLMs handle collaborative information but also present a robust method for enhancing LLM-based recommendation systems.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC