Factor Number Selection in the Tensor Decomposition of EEG Data: Mission (Im)Possible?

A number of factors is an essential parameter in the tensor decomposition methods. It significantly influences not only the decomposition quality but also its interpretation. Many approaches and heuristics were proposed for this purpose. However, their performance is usually demonstrated on data with a simplified structure, and therefore they can produce inferior results when applied to more complex real data. In this study, on a generated dataset closely mimicking the nature of a human multichannel electroencephalogram (EEG), we compared the performance of five methods for selecting the number of factors. We identified the best performing method, but not even this method led to sufficiently acceptable results.