New dropout method makes AI recommenders 2.8x more robust

Researchers from the University of Science and Technology of China propose Sequential Modality Dropout (SMD), a simple yet effective training strategy to make multi-modal sequential recommenders robust to missing modalities (e.g., images or text) at inference time. Unlike traditional recommenders that assume all modalities are always available, real-world product catalogs often have incomplete data. SMD randomly drops entire modality streams during training with probability p, forcing the model to learn to predict the next item without relying on any single modality.

The method was evaluated on four Amazon product domains using four backbone models (MM-SASRec, IISAN, MISSRec, fMRLRec). Under extreme missingness (95% of items lacking text), SMD retained 61% of the full-modality HR@10 accuracy compared to just 22% without SMD—a 2.8x improvement. Text retention improved by 1.0–3.2x across models with minimal impact on full-modality performance. An optional cross-modal reconstruction loss further boosted retention to 98% under severe text missingness. Notably, SMD requires only a four-line change and is architecture-agnostic, making it easy to integrate into existing systems. The paper was accepted at CIKM 2026.