On the effects of data normalization for domain adaptation on EEG data

In Machine Learning (ML), a well-known problem is the Dataset Shift where data in training and test sets can follow different probability distributions, leading ML systems toward poor generalization performances. This intensely felt Brain-Computer Interfaces (BCIs), bio-signals as Electroencephalographic (EEG) are often used. Indeed, EEG signals highly non-stationary both over time between subjects. To overcome this problem, several solutions based on transfer learning approaches such Domain Adaption (DA). cases, however, actual causes of improvements remain ambiguous. paper focuses impact normalization strategies applied together with DA methods. particular, using SEED, DEAP, BCI Competition IV 2a datasets, we experimentally evaluated without It results that choice strategy plays key role classifier performances scenarios, and, often, use only an appropriate schema outperforms technique. For SEED 2a, proper alone cross-subject context allows to reach accuracy 81.52±7.26% 68.52±11.35%, respectively. cross-session context, 86.56±8.15% 67.82±12.48% for be reached, best performance achieved was 39.33±14.08%. All these comparable obtained by strategies.