Distributed non-negative RESCAL with automatic model selection for exascale data

Manish Bhattarai, Namita Kharat, Ismael Boureima, Erik Skau, Benjamin T. Nebgen, Hristo N. Djidjev, Sanjay V. Rajopadhye, James P. Smith, Boian S. Alexandrov

Abstract: Highlights•The first distributed RESCAL implementation to estimate latent features.•pyDRESCALk scales well for extra-large dense and sparse nonnegative tensors.•The first distributed RESCAL framework on distributed GPU/CPU architectures.•We demonstrate pyDRESCALk's ability to decompose 10TB dense and 9EB sparse data.•The library is released for reproducibility and accessibility to researchers.