Web App: MSDA-Bench
Published:
Application: MSDA-Bench Title: Multi-Source Domain Adaptation Benchmark for Cross-Session EEG Classification Author: Yiming Shen License: MIT
Overview
MSDA-Bench is an interactive dashboard for comparing how different source-session utilization strategies handle distribution shift in EEG-based brain-computer interfaces. Built as the visualization companion to the research on Cross-Session BCI Transfer, it provides 10 analysis views across 6 pipelines for systematic evaluation of domain adaptation methods.
Pipelines
| Pipeline | Strategy | Sessions Used |
|---|---|---|
| MAP | Merge All & Predict — uniform pooling of all source sessions | 100% |
| DWP | Distance-Weighted Pooling — soft inverse-distance weighting | 100% |
| MMP_mta | Minimum-distance Multi-source, Merge-Then-Adapt — CI-gated nearest selection + weighted merge | ~40% |
| MMP_moe | Minimum-distance Multi-source, Mixture-of-Experts — CI-gated selection + weighted voting | ~40% |
| BDP_fb | Bridge-Domain Pipeline (far-to-bridge) — hierarchical bridge/far partition with proxy tuning | ~57% |
| BDP_bf | Bridge-Domain Pipeline (bridge-to-far) — reverse proxy direction | ~57% |
Dashboard Pages
- Overview — Dataset summary, completion matrix, QC checks
- Pipeline Benchmark — Head-to-head pipeline comparison with multiple metrics
- Stability & Sensitivity — Config selection premium, ranking stability, variance analysis
- Config Explorer — 24-config heatmap, feature contribution
- Subject Explorer — Single-subject deep-dive
- DA Analysis — Domain adaptation gain/harm analysis
- Mechanism Explorer — Session role visualization (BDP bridge/far, MMP selection, DWP weights)
- Target Session — Accuracy by target session, difficulty ranking
- Prediction Error — Per-class accuracy, confusion matrix, hard sessions
- Efficiency & Progress — Timing, accuracy-time tradeoff
Tech Stack
Python 3.10+, Streamlit, Plotly, pandas, NumPy, SciPy.