You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
|Elhassan & Aljurf | 2016|Class imbalance problem: A review of recent techniques | J. Applied Sciences, 16(8), 314–328 | Comprehensive resampling technique review|
82
+
|Dal Pozzolo et al. | 2015|Calibrating Probability with Undersampling | IEEE SSCI 2015 | Resampling distorts probability estimates|
83
83
| García et al. | 2020 | Understanding the apparent superiority of over-sampling | Expert Systems with Applications, 158, 113026 | Oversampling increases safe minority samples |
84
-
|Hasanin et al. | 2020 | Over- and Under-sampling Approach for EISM Data | Frontiers in Public Health, 8, 178 | HUSDOS-Boost; stratified bagging for EISM |
84
+
|Fujiwara et al. | 2020 | Over- and Under-sampling Approach for EISM Data | Frontiers in Public Health, 8, 178 | HUSDOS-Boost; stratified bagging for EISM |
85
85
| Yang et al. | 2024 | A review on over-sampling techniques for multi-class imbalanced datasets | Frontiers in Digital Health, 6, 1430245 | Multi-class oversampling review |
86
86
| Zhao et al. | 2025 | A Survey on Small Sample Imbalance Problem | arXiv:2504.14800 | EISM definitions; hybrid ensemble recommendations |
87
87
@@ -116,7 +116,7 @@ Based on the literature, Resample Lab identifies three critical regimes:
? `Minority class (N=${minority}) < 50 samples. Standard SMOTE interpolates between sparse noise rather than true structure. Recommend: Hybrid Ensembles (e.g., HUSDOS-Boost) or Stratified Bagging.`
362
362
: `Events Per Variable (EPV=${epv.toFixed(1)}) < 10. Insufficient signal density for reliable synthetic generation.`,
message: 'All resampling methods distort probability estimates. A predicted 80% risk may actually represent 10% true probability. If calibrated probabilities are required, use Class Weights + Isotonic Calibration instead.',
396
-
citationIds: ['elhassan2016']
396
+
citationIds: ['dalpozzolo2015']
397
397
});
398
-
citations.push('elhassan2016');
398
+
citations.push('dalpozzolo2015');
399
399
}
400
400
401
401
// Fold Integrity Warning (always show as best practice)
402
402
regimeWarnings.push({
403
403
type: 'fold-integrity',
404
404
title: 'Cross-Validation Best Practice',
405
405
message: 'Apply Feature Selection and Resampling INSIDE the cross-validation loop, not before. Performing these steps on the full dataset causes data leakage and inflated performance estimates.',
result.rationale=`Resampling would distort probability estimates. Class Weights maintain calibration while addressing imbalance. Apply Isotonic Calibration post-training for optimal Brier Score.`;
result.rationale=`High sparsity (${Math.round(sparsity*100)}%) but concentrated. Standard interpolation (SMOTE) is unreliable. Strategy: Process features separately—interpolate dense columns, impute sparse ones.`;
444
-
result.sparsityWarning="Recommendation: Utilize SMOTE-NC or feature-specific handling.";
result.rationale=`Normally ${dominantOriginal} is best, but ${Math.round(sparsity*100)}% uniform sparsity (Stability: ${stability.toFixed(2)}) renders synthetic neighborhood generation unreliable. Use Class Weights.`;
450
-
result.sparsityWarning="Critical Sparsity: Euclidean distance metrics are unstable.";
result.rationale=`Resampling would distort probability estimates. Class Weights maintain calibration while addressing imbalance. Apply Isotonic Calibration post-training for optimal Brier Score.`;
result.rationale=`High sparsity (${Math.round(sparsity*100)}%) but concentrated. Standard interpolation (SMOTE) is unreliable. Strategy: Process features separately—interpolate dense columns, impute sparse ones.`;
444
+
result.sparsityWarning="Recommendation: Utilize SMOTE-NC or feature-specific handling.";
result.rationale=`Normally ${dominantOriginal} is best, but ${Math.round(sparsity*100)}% uniform sparsity (Stability: ${stability.toFixed(2)}) renders synthetic neighborhood generation unreliable. Use Class Weights.`;
450
+
result.sparsityWarning="Critical Sparsity: Euclidean distance metrics are unstable.";
result.rationale=`Undersampling reduces computational load and noise without disrupting the manifold structure in high dimensions. Consider feature selection to improve distance metric reliability.`;
0 commit comments