The analysis of mobile EEG Brain Computer Interface (BCI) recordings can benefit from unsupervised learning methods. Removing the calibration phase allows for faster and shorter interactions with a BCI and could potentially deal with non-stationarity issues in the signal quality. Here we present a data-driven approach based on a trilinear decomposition, Canonical Polyadic Decomposition (CPD), applied to an auditory BCI dataset. Different ways to construct a data-tensor for this purpose and how the results can be interpreted are explained. We also discuss current limitations in terms of trial identification and model initialization. The results of the new analysis are shown to be comparable to those of the traditional supervised stepwise LDA approach.