SA-RSQ: Sparse Representation Framework for Multi-modal Recommender Systems

_reachsumit · x · 2026-08-25

The paper proposes SA-RSQ (Sparse Activation-based Residual Soft Quantization) to compress multimodal item embeddings. Using Top-K sparse routing and softmax weights, it stores compact tuples to decouple storage from codebook dimensionality, avoiding accuracy loss from hard quantization. Experiments show favorable reconstruction-performance trade-offs at 8-48 bytes/item, with an online A/B test showing a +2.51% CTR lift.

Original post →

More from Research

Research channel →