Boolean Matrix Factorization via Nonnegative Auxiliary Optimization
This work addresses matrix factorization for binary data, which is incremental as it introduces a new optimization approach rather than a paradigm shift.
The paper tackles Boolean matrix factorization by reformulating it as a nonnegative auxiliary optimization problem, proving equivalences and algorithm properties, and demonstrates effectiveness on synthetic and real datasets with comparisons to current methods.
A novel approach to Boolean matrix factorization (BMF) is presented. Instead of solving the BMF problem directly, this approach solves a nonnegative optimization problem with the constraint over an auxiliary matrix whose Boolean structure is identical to the initial Boolean data. Then the solution of the nonnegative auxiliary optimization problem is thresholded to provide a solution for the BMF problem. We provide the proofs for the equivalencies of the two solution spaces under the existence of an exact solution. Moreover, the nonincreasing property of the algorithm is also proven. Experiments on synthetic and real datasets are conducted to show the effectiveness and complexity of the algorithm compared to other current methods.