SymCERE: Symmetric Contrastive Learning for Robust Review-Enhanced Recommendation

TL;DR

SymCERE introduces a unified geometric contrastive learning framework to bridge the Fusion Gap in review-enhanced recommendation. By projecting all embeddings to the unit hypersphere and applying a symmetric NCE that excludes false negatives, the method decouples user preferences from popularity-based magnitudes while aligning graph and text signals. The approach yields substantial improvements across 15 datasets and reveals Semantic Anchoring, where objective product attributes drive robust alignment over generic sentiment. This work advances robust, interpretable multi-modal recommendations by combining principled debiasing with cross-modal, intra-modal, and ranking objectives.

Abstract

Modern recommendation systems fuse user behavior graphs and review texts but often encounter a "Fusion Gap" caused by False Negatives, Popularity Bias, and Signal Ambiguity. We propose SymCERE (Symmetric NCE), a contrastive learning framework bridging this gap via structural geometric alignment. First, we introduce a symmetric NCE loss that leverages full interaction history to exclude false negatives. Second, we integrate L2 normalization to structurally neutralize popularity bias. Experiments on 15 datasets (e-commerce, local reviews, travel) demonstrate that SymCERE outperforms strong baselines, improving NDCG@10 by up to 43.6%. Notably, we validate this on raw reviews, addressing significant noise. Analysis reveals "Semantic Anchoring," where the model aligns on objective vocabulary (e.g., "OEM," "gasket") rather than generic sentiment. This indicates effective alignment stems from extracting factual attributes, offering a path toward robust, interpretable systems. The code is available at https://anonymous.4open.science/r/ReviewGNN-2E1E.

Figures (4)

• Figure 1: Visualizing the Fusion Gap and its Resolution. Left: Before normalization, the embedding space is dominated by magnitude artifacts (popularity bias), causing a geometric and semantic disconnect between Graph (blue) and Text (orange) modalities. Right: SymCERE structurally aligns these spaces on a unit hypersphere, enabling coherent clustering based on semantic preference rather than interaction frequency.
• Figure 2: The SymCERE Framework. The model bridges the Fusion Gap through two key geometric mechanisms: (1) L2 Normalization projects embeddings onto a hypersphere to structurally neutralize popularity bias (magnitude noise). (2) A Symmetric NCE Loss utilizes the known interaction mask $M$ to deterministically exclude false negatives (grey cells) from the denominator, preventing intra-class repulsion. This unified objective aligns collaborative signals ($\mathbf{g}$) and semantic signals ($\mathbf{t}$) while preserving ranking order via BPR.
• Figure 3: Geometric Mechanism of Semantic Anchoring. (a) Standard InfoNCE induces repulsion between semantically similar items (false negatives), causing diffusion on the hypersphere. This allows subjective noise ($\mathbf{v}_{\text{subj}}$) to persist as there is no structural pressure to align. (b) SymCERE removes these repulsive forces via false negative elimination, allowing items to form tight clusters around a common centroid $\boldsymbol{\mu}_{\mathcal{C}}$ aligned with objective features. (c) Semantic Anchoring: Under the strict L2 normalization constraint, maximizing alignment with the centroid forces the embedding to allocate its limited magnitude almost entirely to the objective component ($\lambda_{\text{obj}} \to 1$), effectively suppressing subjective noise ($\lambda_{\text{subj}} \to 0$).
• Figure 4: Impact of L2 Normalization on Dimensional Utilization (Pet Supplies). The orange distribution (With Normalization) exhibits significantly higher variance across dimensions compared to the blue distribution (Without Normalization). This confirms that SymCERE prevents dimensional collapse, utilizing the embedding space more uniformly to encode semantic differences.