Algorithms for Large-Scale Sparse Tensor Factorization

Abstract

Tensor factorization is a technique for analyzing data that features interactions of data along three or more axes, or modes. Many fields such as retail, health analytics, and cybersecurity utilize tensor factorization to gain useful insights and make better decisions. The tensors that arise in these domains are increasingly large, sparse, and high dimensional. Factoring these tensors is computationally expensive, if not infeasible. The ubiquity of multi-core processors and large-scale clusters motivates the development of scalable parallel algorithms to facilitate these computations. However, sparse tensor factorizations often achieve only a small fraction of potential performance due to challenges including data-dependent parallelism and memory accesses, high memory consumption, and frequent fine-grained synchronizations among compute cores. This thesis presents a collection of algorithms for factoring sparse tensors on modern parallel architectures. This work is focused on developing algorithms that are scalable while being memory- and operation-efficient. We address a number of challenges across various forms of tensor factorizations and emphasize results on large, real-world datasets.