Modern recommendation systems can achieve high performance by fusing user behavior graphs (via GNNs) and review texts (via LLMs). However, this fusion faces three significant issues: (1) False Negatives in contrastive learning can degrade the training signal by penalizing similar items; (2) Popularity Bias, often encoded as embedding magnitude, can distort similarity scores; and (3) Signal Ambiguity, which arises from the conflation of objective facts with subjective sentiment in reviews. These interconnected issues can prevent models from learning users' true preferences. In this paper, we propose SymCERE (Symmetric SINCERE), a contrastive learning method that addresses these three issues simultaneously through its structural design. First, we introduce a symmetric application of the SINCERE loss for cross-modal alignment, which is designed to eliminate false negatives in recommendation. Second, by integrating this with L2 normalisation under a "magnitude-as-noise" hypothesis, we aim to mitigate popularity bias by forcing the model to encode preferences primarily in the vector's direction. Experiments on 15 datasets from three distinct platforms (e-commerce, local reviews, and travel) demonstrate that SymCERE outperforms several strong baselines, achieving a relative improvement of up to 43.6% on NDCG@10. Furthermore, a detailed LIME analysis shows that the model learns to anchor alignment on objective, informative vocabulary (e.g., "OEM," "compatible," "gasket"), while placing less emphasis on generic sentiment (e.g., "good," "great"). This suggests that effective semantic alignment stems from understanding factual product attributes, offering a path toward more accurate recommendation systems. The code is available at: https://anonymous.4open.science/r/ReviewGNN-2E1E.
翻译:暂无翻译