Training and Evaluation

This chapter explains how to turn an affective EEG task definition into a defensible experimental protocol. Once the problem formulation is fixed, the next step is to decide how data should be split, how leakage should be prevented, what metrics should be reported, and what kind of evidence is actually needed to support a performance claim.

In EEG-based affective computing, evaluation details often matter as much as the model itself. Random window shuffling, normalization across the full dataset, or misuse of future context can make results look stronger than they really are. For that reason, this chapter emphasizes train-test splits and evaluation protocols that preserve the intended generalization challenge.

The discussion in this chapter follows the task taxonomy introduced earlier. It first defines leakage-safe splits and protocols for batch, online, within-subject, cross-session, chronological, and subject-independent settings. It then covers supervised, unsupervised, self-supervised, representation-learning, and reinforcement-learning training paradigms; training, validation, model selection, calibration, data selection, and EEG-safe augmentation; metrics, statistical evidence, robustness analysis, reproducibility; qualitative assessment of whether a model's behavior is consistent with psychological, clinical, causal, and geometric expectations; transfer learning with few-shot and zero-shot generalization, which are particularly important when labeled data from the target subject, session, or device are scarce; and useful platforms and tools for implementing, benchmarking, and documenting these workflows. Each component should match the intended deployment scenario rather than merely maximize a benchmark score.