nimfa

Why we included this project

Gene expression counts, image pixel intensities, document-term matrices, and audio spectrograms are all naturally nonnegative, and that is the kind of data Nimfa is built for. The library implements many nonnegative matrix factorization algorithms, from alternating least squares and projected gradient methods to more specialized variants, along with initialization strategies and quality metrics like residual sum of squares and sparseness. It handles both dense and sparse matrices, so large inputs that are mostly zeros are not a problem. Researchers and analysts who want to compare factorization approaches on their own data get all of these methods behind one consistent Python API, and the package installs cleanly through conda-forge. The project started in 2011 as a Google Summer of Code effort and has been maintained by volunteers since. It remains a solid choice for classical latent-feature extraction rather than a neural-network-based embedding tool.