Parallel exploration of partial solutions in Boolean matrix factorization

Boolean matrix factorization (BMF) is a well established method for preprocessing and analysis of data. There is a number of algorithms for BMF, but none of them uses benefits of parallelization. This is mainly due to the fact that many of the algorithms utilize greedy heuristics that are inherently sequential. In this work, we propose a general parallelization scheme for BMF in which several locally optimal partial matrix decompositions are constructed simultaneously in parallel, instead of just one in a sequential algorithm. As a result of the computation, either the single best final decomposition or several top-k of them may be returned. The scheme can be applied to any sequential heuristic BMF algorithm and we show the application on two representative algorithms, namely GRECoND and Asso. Improvements in decompositions are presented via results from experiments with the new algorithms on synthetic and real datasets. (C) 2018 Elsevier Inc. All rights reserved. .