Cross-Subject Event-Related Potential Classification via Multi-View Based Contrastive Learning

Abstract

Background:

Event-related potentials (ERPs) provide implicit feedback and error-correction signals that are valuable for brain–computer interfaces (BCIs). However, models trained on source-domain subject data are vulnerable to inter-subject variability and acquisition noise, which substantially degrades generalization to unseen subjects.

Objective:

We propose a multi-view contrastive learning domain generalization (MVCLDG) method to improve cross-subject generalization in ERP recognition by jointly exploiting discriminative feature extraction and domain-invariant representation learning.

Methods:

MVCLDG employs a multi-view feature-extraction module that fuses raw electroencephalography with phase information derived from the Hilbert transform via multi-scale inception blocks, thereby capturing both amplitude and phase features. The model then applies domain-alignment and contrastive-learning constraints to reduce distributional discrepancy across domains, compact within-class representations, and enlarge between-class separability. The approach was evaluated on a public Error-Related Negativity (ERN) dataset and a self-collected semantic–syntactic violation dataset; performance was assessed in cross-subject settings, and ablation and visualization analyses were conducted to probe the contributions of components and neurophysiological interpretability.

Results:

MVCLDG outperformed baseline and representative domain generalization methods in cross-subject ERP recognition without requiring additional target-domain adaptation. Ablation experiments confirmed the effectiveness of each component. Eigen-Class Activation Maps visualizations indicate consistency between the model-attended electrodes and known neurophysiological scalp patterns, supporting both the model’s generalization mechanism and its biological interpretability.

Conclusions:

MVCLDG offers an effective strategy for integrating phase-aware multi-view feature mining with contrastive domain generalization, yielding improved and interpretable cross-subject ERP recognition. The method advances the feasibility of ERP-based closed-loop BCIs that generalize across users.

Impact Statement

This study introduces a novel multi-view contrastive learning domain generalization (MVCLDG) framework. By integrating raw electroencephalography (EEG) with Hilbert-transformed EEG as complementary views, MVCLDG captures joint amplitude–phase features and, via multi-view contrastive learning, extracts domain-invariant representations. Together, these mechanisms improve cross-subject generalization by integrating discriminative feature mining with domain-invariant representation learning. In addition, we constructed a Chinese semantic–syntactic violation event-related potential (ERP) dataset, thereby addressing a critical gap in language-related ERP resources. This work not only provides new directions for improving cross-subject ERP recognition but also lays the groundwork for generalizing ERP-based closed-loop brain–computer interfaces across users.