Sequential nonnegative tucker decomposition on multi-way array of time-frequency transformed event-related potentials

Abstract

Tensor factorization has exciting advantages to analyze EEG for simultaneously exploiting its information in the time, frequency and spatial domains as well as for sufficiently visualizing data in different domains concurrently. Event-related potentials (ERPs) are usually investigated by the group-level analysis, for which tensor factorization can be used. However, sizes of a tensor including time-frequency representation of ERPs of multiple channels of multiple participants can be immense. It is time-consuming to decompose such a tensor. The low-rank approximation based sequential nonnegative Tucker decomposition (LraSNTD) has been recently developed and shown to be computationally efficient with respect to some benchmark datasets. Here, LraSNTD is applied to decompose a fourth-order tensor representation of ERPs. We find that the decomposed results from LraSNTD and a benchmark nonnegative Tucker decomposition algorithm are very similar. So, LraSNTD is promising for ERP studies.