+
🔍 Shortcut Detection Report
+
+
+
+
🟡 MODERATE RISK: One method detected shortcuts
- • Geometric analysis signaled moderate shortcut risk
+
+
📊 Visualizations
+
+
PCA Projection
+
t-SNE Projection
+
+
3D Embedding Analysis
+
+Interactive 3D
+
+Static view (if interactive does not load):
+
+
+
Statistical Test P-values
Top 20 most significant dimensions shown by default.
+
Expand to full heatmap
+
+
+
+
+
🧭 Geometric Analysis
+
Risk: Moderate
+
Reason: One geometric test indicates shortcut risk.
+
One geometric test indicates shortcut risk.
+
Top Bias Direction Pairs
+
| Groups | Effect Size | Projection Gap | Alignment |
|---|
+0 vs 1 | 0.454 | 0.167 | 0.623 |
+
+
Top Subspace Overlap Pairs
+
| Groups | Mean Cosine | Min Angle (°) | Max Angle (°) |
|---|
+0 vs 1 | 0.951 | 6.0 | 33.6 |
+
+
+
+
🔬 Probe-based Detection
+
Risk: Low
+
Reason: f1 (0.529) compared against threshold (0.700).
+
f1: 52.94%
+
+
+
Embedding Frequency Shortcut
+
Risk: Low
+
Reason: 0/2 classes exceeded TPR/FPR thresholds (0.500/0.150).
+
Probe accuracy: 67.00%
+
Shortcut classes: []
+
Top-percent used: 0.05
+
Class Rates
+
+| Class | TPR | FPR | Support | Top Dims |
|---|
+0 | 0.663 | 0.323 | 101 | [] |
+1 | 0.677 | 0.337 | 99 | [28, 18] |
+
+
+
+
📈 Statistical Testing
+
Risk: Low
+
Reason: 0/1 comparisons have significant features after FDR_BH (alpha=0.050).
+
Group Comparisons
+
+| Comparison | Significant Features |
|---|
+0_vs_1 | 0 |
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
💡 Recommendations
+
No strong shortcuts detected. Continue monitoring with larger/different datasets.
+
General advice: Always validate findings with domain experts and test mitigation strategies on held-out data.
+
+
+
+
+
