When Modalities Go Missing: Early Fusion for Multimodal Recommendation

Existing multimodal recommenders largely rely on late fusion, where modalities are separately encoded and aligned before fusion. Such designs work well with complete content but become fragile in practice: modalities are often missing, alignment breaks down, and imputation adds complexity with limited benefit. We propose EFMRec, a missing-aware early-fusion framework that follows a relaxed principle of leveraging modalities when available and ignoring them when absent. By exploiting pretrained multimodal models, EFMRec projects available modalities into a shared semantic space and aggregates them through a missing-aware early fusion, producing unified representations without reconstruction or auxiliary losses. These representations are further propagated via a GCN-based architecture and integrated with a collaborative filtering backbone to jointly model multimodal and collaborative signals. Experiments on three benchmarks demonstrate that EFMRec consistently outperforms strong baselines under both full- and missing-modality settings, highlighting its robustness to modality incompleteness.