HASI is an algorithm for low-rank matrix completion described in reference [1].

It uses nonconvex nuclear penalties arising from a hierarchical sparsity inducing prior on singular values. The algorithm iteratively performs adaptive weighted soft thresholded SVD.

Applications are in Collaborative Filtering (predicting user preferences for items), image inpainting, imputation of missing values, etc.

The software is distributed as a Matlab package. It makes use of the PROPACK algorithm for handling large scale matrices.

1. Download and extract HASI.
2. Add folders Matlab_files and PROPACK_utils to Matlab path.
3. Run install_mex.m to install mexfiles.
4. See and run demo_hasi.m.

• ha_soft_impute: the main function that runs HASI algorithm (see [1]).

We also provide:

• soft_impute: runs Soft-impute algorithm (see [2]), special case of HASI with gamma variant and infinite beta parameter.
• hard_impute: runs Hard-impute algorithm (see [2]).
• spectral_norm: computes the largest singular value of a sparse matrix.

Any function help is available via the command help funcname.

HASI software was written by Adrien Todeschini ([email protected]).

HASI software is adapted from the Soft-Impute Matlab code written by Rahul Mazumder with considerable input from Trevor Hastie based on reference [2].

[1]: "Probabilistic Low-Rank Matrix Completion with Adaptive Spectral Regularization Algorithms" by Adrien Todeschini, François Caron, Marie Chavent (NIPS' 2013)

[2]: "Spectral Regularization Algorithms for Learning Large Incomplete Matrices" by Rahul Mazumder, Trevor Hastie, Rob Tibshirani (JMLR vol 11, 2010)

• fix binary case

Original release