Proceedings of the Annual Meeting of the Cognitive Science Society

In stimulus-driven EEG emotion classification, labels are defined at the trial level. Each trial corresponds to a fixed stimulus, and windowing yields multiple dependent repeated measurements within the same trial. However, standard pipelines treat windows as independent samples, which ignores the trial as the true unit and encourages models to latch onto transient noise, leading to inconsistent within-trial predictions and unstable generalization. We turn the trial-as-unit prior in stimulus-driven emotion EEG into a trial-defined structural constraint. Specifically, windows are grouped by their trial IDs and a loss is added that makes predictions for windows within the same trial agree with each other. This penalizes reliance on window-specific transient noise and encourages the model to encode stable, trial-level representations. Across SEED and SEED-IV and three backbones including EEGNet, EEGConformer, and CTNet, the proposed constraint increases within-trial prediction consistency and stabilizes validation behavior. It also improves cross-session classification accuracy.