Constructing a Personalized Cross-Day EEG-Based Emotion-Classification Model Using Transfer Learning.

State-of-the-art electroencephalogram (EEG)-based emotion-classification works indicate that a personalized model may not be well exploited until sufficient labeled data are available, given a substantial EEG non-stationarity over days. However, it is impractical to impose a labor-intensive, time-consuming multiple-day data collection. This study proposes a robust principal component analysis (RPCA)-embedded transfer learning (TL) to generate a personalized cross-day model with less labeled data, while obviating intra- and inter-individual differences. Upon the add-session-in validation on two datasets MDME (five-day data of 12 subjects) and SDMN (single-day data of 26 subjects), the experimental results showed that TL enabled the classifier of an MDME individual (using his/her 1st-day session only) to improve progressively in valence and arousal classification by adding similar source sessions (SSs) via the within-dataset TL (wdTL) and cross-dataset TL (cdTL) manners. When recruiting three SSs to test on the 5th-day session, the wdTL improvement (valence: 11.19%, arousal: 5.82%) marginally outperformed the subject-dependent (SD) counterpart (valence: 9.75%, arousal: 3.77%) that was obtained using their own 2nd-4th-day sessions only. The cdTL returned a similar trend in valence (8.35%), yet it was less effective in arousal (0.81%). Most importantly, such cross-day enhancements did not occur in either SD or TL scenarios until RPCA processing. This work sheds light on how to construct a personalized model by leveraging ever-growing EEG repositories.