Consistency-guided evidential debiasing for multimodal recommendation

Multimodal recommendation integrates auxiliary modalities, such as images and text, to enrich item representations and improve personalized ranking. However, recommendation accuracy can be affected by cross-modal bias, where single-modality signals may dominate multimodal representations and lead to unreliable recommendations. Handling this bias introduces two challenges: (1) Conventional multimodal fusion strategies may dilute misleading single-modality signals during aggregation, resulting in insufficient debiasing penalties. (2) Debiasing without evaluating cross-modal semantic consistency may misidentify semantically consistent items and impose unnecessary penalties. To address these issues, we propose a Consistency-guided Evidential Debiasing framework for multimodal Recommendation (CEDRec). Specifically, CEDRec utilizes an evidential bias network that adopts a simple Dempster-Shafer formulation to represent deceptive evidence and uncertainty for each modality. Subsequently, the network performs evidential fusion over these modality-specific estimates to obtain a reliable cross-modal bias score while alleviating the dilution of strong unimodal signals. Moreover, to prevent misidentifying semantically consistent items, CEDRec employs a semantic bottleneck network based on the variational information bottleneck principle. This network distills shared cross-modal semantics to generate a dynamic gating signal, thus protecting semantically consistent items from unnecessary penalties. Finally, during the joint optimization, gradient projection is employed to resolve conflicts between collaborative preference learning and semantic bottleneck compression, ensuring that shared semantics are distilled without sacrificing ranking performance. Extensive experiments on multiple real-world datasets demonstrate that CEDRec effectively mitigates cross-modal bias and achieves a more balanced item exposure distribution while maintaining competitive ranking performance.