Distributed non-negative RESCAL with Automatic Model Selection for Exascale Data

With the boom in development of computer hardware and software, social media, IoT platforms, communications, there has been exponential growth volume data produced worldwide. Among these data, relational datasets are growing popularity as they provide unique insights regarding evolution communities their interactions. Relational naturally non-negative, sparse, extra-large. usually contain triples (subject, relation, object) represented graphs/multigraphs, called knowledge graphs, which need to be embedded into a low-dimensional dense vector space. various embedding models, RESCAL allows learning extract posterior distributions over latent variables make predictions missing relations. However, is computationally demanding requires fast distributed implementation analyze extra-large real-world datasets. Here we introduce non-negative algorithm for heterogeneous CPU/GPU architectures with automatic selection number (model selection), pyDRESCALk. We demonstrate correctness pyDRESCALk large synthetic tensors efficacy showing near-linear scaling that concurs theoretical complexities. Finally, determines an 11-terabyte 9-exabyte sparse tensor.