P300-Based Partial Face Recognition With xDAWN Spatial Filter and Covariance Matrix

Face cognition is one of the most crucial cognition processes in social interaction. In the study of face cognition, rapid serial face cognition (RSVP), the presentation of target and non-target images, is often used to understand the cognition mechanism. When a person perceives the target image, the event-related potential (ERP) is evoked. To identify the target image or the event of interest of a person, the classification model machine learning is introduced. However, the machine learning model that works the best when applied to ERP is still in question. This study aimed to investigate the simplest machine learning model that performs best when comparing six classification models applied to ERP peak evoked during the partial face cognition task. The six models used in this investigation were linear discrimination analysis (LDA), xDAWN filter + linear support vector machine (SVM), xDAWN filter + LightGBM, xDAWN covariance matrix + tangent space + linear SVM, xDAWN covariance matrix + tangent space + LightGBM, and xDAWN covariance matrix + minimum distance to mean (MDM). As a result, we found that the xDAWN covariance matrix improved the classification performance compared to combining the xDAWN filter with the same classification models. In addition, the combination of the xDAWN covariance matrix and MDM provided the best performance in participant-dependent cross-validation. In contrast, the xDAWN covariance matrix, tangent space, and LightGBM provided the most promising performance in the participant-independent cross-validation.