Scholarlink

While EEG-based affective computing holds significant promise, its progress is currently stifled by the lack of standardization across datasets—specifically regarding incompatible electrode configurations and inherent subject variability—which severely restricts model transferability. To address this, we present a universal pre-training framework that introduces a two-phase learning strategy. (1) Channel-independent pre-training: establishes temporal representations via contrastive learning on single channels. This phase utilizes our Unified Channel Schema (UCS) to construct a superset of all sensor positions, ensuring full data utilization despite varying montages (e.g. DEAP-32ch,SEED-62ch). (2) Spatial-temporal fine-tuning: employs a hybrid GCN-Transformer architecture to capture inter-channel dependencies for downstream emotion tasks. Extensive experiments demonstrate significant performance gains over training from scratch on SEED (+30.56%), DEAP (+20.73%), and the DREAMER (+10.15%) dataset. Our framework achieves state-of-the-art 97.50% on SEED and 77.86% on DEAP. The UCS proves to be a scalable solution for integrating heterogeneous data, establishing a solid foundation for future general-purpose EEG modeling.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

From Heterogeneity to Harmony: Universal Pretraining for EEG-Based Emotion Recognition

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • You Li,
  • Hao Wang,
  • Xiang Li

摘要

While EEG-based affective computing holds significant promise, its progress is currently stifled by the lack of standardization across datasets—specifically regarding incompatible electrode configurations and inherent subject variability—which severely restricts model transferability. To address this, we present a universal pre-training framework that introduces a two-phase learning strategy. (1) Channel-independent pre-training: establishes temporal representations via contrastive learning on single channels. This phase utilizes our Unified Channel Schema (UCS) to construct a superset of all sensor positions, ensuring full data utilization despite varying montages (e.g. DEAP-32ch,SEED-62ch). (2) Spatial-temporal fine-tuning: employs a hybrid GCN-Transformer architecture to capture inter-channel dependencies for downstream emotion tasks. Extensive experiments demonstrate significant performance gains over training from scratch on SEED (+30.56%), DEAP (+20.73%), and the DREAMER (+10.15%) dataset. Our framework achieves state-of-the-art 97.50% on SEED and 77.86% on DEAP. The UCS proves to be a scalable solution for integrating heterogeneous data, establishing a solid foundation for future general-purpose EEG modeling.