Distributed Nonnegative Matrix Factorization with HALS Algorithm on Apache Spark

Abstract

Nonnegative Matrix Factorization (NMF) is a commonly-used unsupervised learning method for extracting parts-based features and dimensionality reduction from nonnegative data. Many computational algorithms exist for updating the latent nonnegative factors in NMF. In this study, we propose an extension of the Hierarchical Alternating Least Squares (HALS) algorithm to a distributed version using the state-of-the-art framework - Apache Spark. Spark gains its popularity among other distributed computational frameworks because of its in-memory approach which works much faster than well-known Apache Hadoop. The scalability and efficiency of the proposed algorithm is confirmed in the numerical experiments, performed on real data as well as synthetic ones.