Dae Hyeon Kim RESEARCH 한국어

Multimodal data and domain adaptation

Evaluated multimodal fusion and adversarial domain adaptation under subject and session changes.

GitHub: MDA-GCL
Role
Research contributor; co-first author of MDA-GCL
Contribution
Co-developed the domain-adaptation model and ablations. Contributed to multimodal acquisition protocol design.
Period & affiliation
2026.04 ~ Present IITP

01 / RESEARCH DETAILMDA-GCL: multimodal fusion and domain alignment

Challenge
EEG and eye-feature distributions change across subjects and sessions.
Approach
Applied graph contrastive and domain-adversarial objectives to a graph built through modality-similarity fusion.
Contribution
Co-developed the model and ablations as a co-first author and contributed to acquisition protocols.
Modality similarity graphs, contrastive learning and domain-adversarial learning
Modality similarity graphs, contrastive learning and domain-adversarial learning
View full-size figure
SEED-IV cross-subject evaluation
SEED-IV cross-subject evaluation
View full-size figure
DANN ablation (%) Accuracy: Without DANN 62.5, Full model 76.1 0 23 46 70 93 62.5 Without DANN 76.1 Full model Accuracy
DANN ablation (%)
Domain-adaptation accuracy (%) and acquisition target
Accuracy (%) Cross-subject Cross-session
SEED-IV 76.1 ± 7.4 83.8 ± 8.3
SEED-V 75.6 ± 6.3 82.5 ± 9.1
Evaluation conditions and scopeTransductive domain adaptation uses unlabeled target features. Target emotion labels are excluded from training and model selection. SEED-IV cross-subject/session accuracy: 76.1 ± 7.4% / 83.8 ± 8.3%.