Quantile Multiplicative Updates for Corruption-Robust Nonnegative Matrix Factorization

Zane Collins, Jamie Haddock, Tyler Headley, Luke Wang

Session: General

Keywords: Nonnegative matrix factorization, corruption-robust model, quantile-based methods, multiplicative updates

TL;DR: We introduce a quantile-based variant of the popular multiplicative updates method for training the Frobenius norm-formulation of NMF which avoids the effects of corruption in the data.

Abstract: Nonnegative matrix factorization (NMF) models have found success in a variety of applications, including document clustering and classification, image processing, and bioinformatics. However, the optimization formulations typically employed for NMF models are often very sensitive to noise and corruption in the data. We introduce a quantile-based variant of the popular multiplicative updates method for training the Frobenius norm-formulation of NMF which avoids the effects of corruption in the data. Our numerical experiments illustrate the promise of this method, and shows that in some scenarios this method applied to the corrupted data recovers factorizations nearly as good as factorizations learned on the uncorrupted data.

Submission Number: 84