Feature Selection

Choosing a subset of features (channels, frequency bands, time points, components) to reduce dimensionality, improve generalization, lower computation, and aid interpretation. Methods are filters (rank features by a univariate score such as mutual information, Fisher score, or r^2 with the label, independent of the classifier), wrappers (search subsets by cross-validated classifier performance — powerful but expensive and leakage-prone), and embedded methods (selection built into training, such as L1/lasso sparsity or the automatic band and CSP selection in FBCSP).

In BCI, selection both helps (fewer channels mean faster setup and less overfitting) and endangers validity: any selection that consults the labels must be redone inside every cross-validation fold, or the reported accuracy is inflated by leakage. Stability of the selected set across folds and subjects is itself a useful diagnostic — features that survive resampling are more trustworthy than a single-fold ranking.