Graph Neural Networks (GNNs) have shown great promise in recommendation systems but struggle with the item cold-start problem due to sparse interactions. A prevailing solution is to leverage multi-modal side information (e.g., images and text) to enhance item representations. However, our empirical analysis reveals a paradox: existing state-of-the-art models often perform worse when fusing multiple modalities compared to utilizing a single modality. We argue that this "more is less" phenomenon stems from two fundamental limitations: the deterministic nature of existing fusion strategies, which neglects epistemic uncertainty and allows noisy modalities to degrade the final representation, and the vulnerability of encoders to spurious correlations. To mitigate this paradox, we propose a novel framework named Variational Counterfactual GNN (VC-GNN). First, we introduce a Variational Semantic Graph Encoder that models item representations as Gaussian distributions rather than deterministic vectors. This allows the model to explicitly quantify the uncertainty of each modality. Second, we design an Uncertainty-Aware Fusion Mechanism based on the Product-of-Experts principle, which dynamically prioritizes reliable modalities while suppressing uncertain ones. Finally, to ensure robustness against spurious correlations, we incorporate a Counterfactual Contrastive Learning module. By generating simulated counterfactual samples, we force the model to ignore spurious shortcuts and focus on intrinsic item features. Extensive experiments on three benchmark datasets demonstrate that VC-GNN outperforms state-of-the-art baselines and mitigates the multi-modal paradox.
Zou et al. (2026) studied this question.