New AI method shrinks recommendation embeddings by 100x with minimal accuracy loss

Researchers have developed a training strategy that replaces dense embedding layers with high-dimensional sparse ones for recommendation systems. By modifying the production-grade ELSA autoencoder, they achieved a 10x reduction in embedding size with no accuracy loss, and a 100x reduction with only a 2.5% performance drop. The sparse embeddings also create an interpretable inverted-index structure that segments items, enabling new functionality like 2D homepage layouts directly within the retrieval model.