Research map / Label-efficient and robust learning / Recognition and long-tail
DMMR: Cross-Subject Domain Generalization for EEG-Based Emotion Recognition via Denoising Mixed Mutual Reconstruction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DMMR, a two-stage pre-train/fine-tune LSTM multi-decoder mutual-reconstruction model with hidden-layer mixup and Time Steps Shuffling noise injection for cross-subject EEG emotion recognition under leave-one-subject-out domain generalization.
Problem
Cross-subject domain generalization for EEG-based emotion recognition: models trained on source subjects must generalize to unseen target subjects despite individual variability, scarce source data, and noise.
Approach
A shared LSTM encoder with subject-specific decoders reconstructs same-emotion features across subjects (mutual reconstruction), with a hidden-layer mixup of decoder outputs for augmentation and a Time Steps Shuffling noise injection that keeps the last time step fixed; adversarial subject discrimination (DG-DANN) is applied in pre-training and then an emotion classifier is fine-tuned.
Key results
- DMMR reaches 88.27% (±5.62) on SEED under leave-one-subject-out, the best among compared methods (GMSS 86.52, Clisa 86.40).
- DMMR reaches 72.70% (±8.01) on SEED-IV, slightly below GMSS (73.48%) but above RGNN (71.65%).
- Ablation: removing noise injection raises SEED-IV accuracy slightly (72.93) but greatly increases std (12.16), and removing mix lowers SEED to 85.42.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add mixed data augmentation (ablation) | DMMR · SEED | accuracy: 85.42 -> 88.27 | Helped |
| add mixed data augmentation (ablation) | DMMR · SEED-IV | accuracy: 71.12 -> 72.70 | Helped |
| add noise injection (ablation) | DMMR · SEED | accuracy: 87.15 -> 88.27 | Helped |
| add noise injection (ablation) | DMMR · SEED-IV | accuracy: 72.93 -> 72.70 | Hurt |
| add noise injection (ablation) | DMMR · SEED-IV | standard deviation: 12.16 -> 8.01 | Helped |
| replace Time Steps Shuffling with Mask Time Steps (ablation) | DMMR · SEED and SEED-IV | accuracy: 88.27 -> 85.46; 72.70 -> 70.83 | Hurt |
| replace Time Steps Shuffling with Channels Shuffling (ablation) | DMMR · SEED and SEED-IV | accuracy: 88.27 -> 86.83; 72.70 -> 71.65 | Hurt |
| replace Time Steps Shuffling with Mask Channels (ablation) | DMMR · SEED and SEED-IV | accuracy: 88.27 -> 86.54; 72.70 -> 72.41 | Hurt |
| replace Time Steps Shuffling with Dropout (ablation) | DMMR · SEED and SEED-IV | accuracy: 88.27 -> 86.37; 72.70 -> 71.21 | Hurt |
Surprising findings
- Removing the noise injection module slightly increased SEED-IV mean accuracy (72.93 vs 72.70) while raising std from 8.01 to 12.16.
- Four of the alternative noise injection methods (mask time steps, channel shuffling, mask channels, dropout) performed worse than no noise injection on SEED.
Limitations
- SEED-IV result is slightly below GMSS, which the authors attribute to the smaller dataset limiting the augmentation.
- Relies on source-subject labels for the reconstruction loss, so requires labeled source data.
- Evaluated only on SEED and SEED-IV with DE features from the first session.
Experimental setup
Datasets: SEED SEED-IV
Benchmarks: SEED SEED-IV
Metrics: accuracy (avg) standard deviation
Baselines: DGCNN BiHDM RGNN GMSS DG-DANN DResNet Clisa PPDA-NC PPDA
Builds on: MTAE DiMAE PPDA DG-DANN mixup
Closest papers by meaning
- Multi-dataset Joint Pre-training of Emotional EEG Enables Generalizable Affective Computing (NeurIPS 2025 · similarity 0.90)
- Beyond Mimicking Under-Represented Emotions: Deep Data Augmentation with Emotional Subspace Constraints for EEG-Based Emotion Recognition (AAAI 2024 · similarity 0.86)
- FEEL: Quantifying Heterogeneity in Physiological Signals for Generalizable Emotion Recognition (NeurIPS 2025 · similarity 0.86)
- Multi-to-Single: Reducing Multimodal Dependency in Emotion Recognition Through Contrastive Learning (AAAI 2025 · similarity 0.85)
- What do EEG Foundation Models Capture from Human Brain Signals? (NeurIPS 2026 · similarity 0.85)
- CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.