ASTEN: An accurate and scalable approach to Coupled Tensor Factorization

Abstract

Coupled Tensor Factorization (CTF) has become one of the most popular methods for joint analysis of high dimensional data generated from multiple sources. The goal of CTF is to factorize correlated datasets into latent factors efficiently. This research was taken with a particular goal of improving the accuracy of CTF. It is important to optimize the factorization of each single tensor of the coupled tensors. To achieve this, we introduce ASTEN, an Accurate and Scalable Tensor factorization method, where the objective function is optimized with respect to every single tensor and matrix. Differing from algorithms with a traditional objective function which forces shared modes among tensors to have identical factors, ASTEN enables each tensor to have its own discriminative factor on the shared mode and thus is capable of finding the accurate approximation of every tensor. Furthermore, to make it highly scalable in handling big data, we design it to be fully distributed and scalable with respect to the number of tensors, their dimensions, their sizes and the number of data partitions. In addition, we provide our theoretical proof and experimental evidence that our algorithm converges to an optimum. Experiments on both real and synthetic datasets demonstrate that our proposed ASTEN outperforms alternative existing algorithms.