Skip to content

recombee/CompresSAE

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

2 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Embedding Compression for Scalable Retrieval in Recommender Systems

arXiv By Recombee Research

Recommender systems embeddings are growing. Sparsity is here to help.

Results Our learnable sparse compression algorithm, CompresSAE, achieves a superior compression-retrieval accuracy trade-off, outperforming equally sized Matryoshka embeddings and approaching uncompressed embedding performance with 12× fewer parameters.

Model Architecture (CompresSAE)

CompresSAE Architecture CompresSAE is a sparse autoencoder (SAE) that maps dense embeddings into high-dimensional, sparsely activated vectors optimized for fast similarity search.

Two inference modes allow a trade-off between latency and accuracy: fast retrieval computes similarity in the sparse compressed space, while high-accuracy retrieval uses similarity in the dense reconstructed space.

See model.py for implementation.

License

MIT License

About

Sparse Embedding Compression for Scalable Retrieval in Recommender Systems

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 2

  •  
  •  

Languages