⚖️ Imbalanced Learning Workshop

Handle rare events with resampling (undersample, oversample, SMOTE), class weighting, focal loss, and threshold tuning. Inspect PR/ROC curves and the confusion matrix.

Dataset & Model PR / ROC Controls Guide
Beginner Quick Start
  1. Generate data: Set Total samples, Pos % (minority prevalence), and Noise, then click Generate.
  2. Choose resampling (optional): Undersample/Oversample/SMOTE to balance the training set, then click Apply Resampling.
  3. Train: Use Train Step or Train 100. Increase Class weight (pos) or Focal γ if the model ignores positives.
  4. Decide: Move the Threshold slider to trade precision vs recall. Watch the confusion matrix and PR/ROC curves update.
Tips: PR curve (left) is more informative for imbalance. The colored background shows predicted probability; the dashed line is the decision boundary.
Dataset & Model
Positive (1) Negative (0)
Tip: Click the canvas to add a point to the selected class. Apply resampling to adjust training distribution.
Precision
—
Recall
—
F1
—
PR AUC
—
Pred 0Pred 1
Actual 0——
Actual 1——
Threshold0.50
Adjust decision threshold to trade precision vs recall.
Precision–Recall & ROC
PR curve is informative for imbalanced data.
Diagonal line is random classifier.
📦 Dataset
Pos % controls the rarity of the positive class. Higher Noise makes classes overlap and the task harder.
🔁 Resampling
Undersample = faster but may drop information. Oversample = simple duplication. SMOTE = creates synthetic minority points between neighbors.
Target Pos %40
Resampling affects training distribution only; evaluation uses all points.
🤖 Model (Logistic)
LR0.10
Epochs/Step10
Class weight (pos)3.0
Focal γ0.0
Class weight (pos) increases penalty for missing positives (↑ recall). Focal γ focuses learning on hard examples (try 1.5–2.0 for heavy imbalance).
About Imbalanced Learning
Why it’s hard
  • Minority class is rare → models bias to predicting majority
  • Accuracy can be misleading → prefer PR curves and recall
  • Threshold choice matters as much as model choice
Techniques
  • Resampling: Undersample majority, oversample minority, or use SMOTE to synthesize new minority samples
  • Class weights: Penalize mistakes on minority more heavily
  • Focal loss: Down-weight easy examples to focus on hard ones
How to evaluate
  • PR curve & AP: Better for imbalance than ROC
  • Confusion matrix at threshold: Inspect precision/recall trade-offs
  • Calibration: If making probabilistic decisions, calibrate then choose thresholds
Workflow tips
  • Try resampling + class weights, then tune threshold by business costs
  • Use focal loss if minority still ignored
  • Validate with PR AUC and recall at target precision
How it works and how to visualize effects
How it works (pipeline)
  1. Generate data: Set total size, positive percentage, and noise; points are drawn from two clusters.
  2. Choose resampling: Applies only to the training indices (display shows all points). Options:
    • Undersample: reduce majority
    • Oversample: replicate minority
    • SMOTE: synthesize minority near neighbors
  3. Train model: Logistic regression updates weights with:
    • Class weight (pos): upweights minority errors
    • Focal γ: emphasizes hard examples (down-weights easy ones)
    • LR / Epochs: controls update magnitude and iterations
  4. Decide: Threshold converts probabilities to classes → confusion matrix and metrics update live.
  5. Evaluate: PR/ROC curves recomputed from model probabilities over all points (not just resampled subset).
Tip: Resampling shifts training distribution; evaluation remains on the full set to avoid inflated metrics.
How to see and visualize improvements
  • Baseline vs Resampling: Generate with low Pos% (e.g., 10–15%). Train baseline (no resampling). Apply SMOTE or Oversample and retrain. Watch recall and PR AUC rise.
  • Class weights: Increase “Class weight (pos)” and observe the decision boundary tilt toward the minority and recall improve (precision may drop).
  • Focal loss: Set γ=2.0 for highly imbalanced/noisy data. Train; hard minority points influence more—recall improves.
  • Threshold tuning: Move the threshold slider:
    • Lower threshold → higher recall, more FP
    • Higher threshold → higher precision, more FN
  • PR vs ROC: Use PR to compare scenarios; it’s more sensitive to improvements in minority detection than ROC on imbalanced sets.
  • Stress test: Increase Noise (σ) to 0.9–1.2; compare “None” vs “SMOTE + class weights” for robustness.
  • Visual boundary: The colored background shows predicted probability; the dashed line is the current decision boundary.
Suggested experiments
  • Pos% = 5–10%, SMOTE to 40% target → train → compare PR AUC and recall
  • Keep data fixed; sweep class weight from 1.0 → 5.0 and plot F1 changes
  • Combine Oversample + γ=1.5 and compare to SMOTE-only

Support This Free Tool

Every coffee helps keep the servers running. Every book sale funds the next tool I'm dreaming up. You're not just supporting a site — you're helping me build what developers actually need.

500K+ users
200+ tools
100% private
☕ One-time support Buy me a coffee 📚 Learn & support 9-Book Bundle - $9 Stay updated Follow @anish2good
Privacy Guarantee: Private keys you enter or generate are never stored on our servers. All tools are served over HTTPS.

About This Tool & Methodology

This workshop explores class imbalance strategies: resampling (undersample/oversample), class weights, and threshold tuning. It visualizes confusion matrices, ROC/PR curves, and metrics sensitive to imbalance (F1, MCC).

Learning Outcomes

  • Understand why accuracy can be misleading under imbalance.
  • Compare class weighting vs sampling; observe PR curve behavior.
  • Use MCC/F1/recall@precision as robust alternatives.

Authorship & Review

  • Author: 8gwifi.org engineering team
  • Reviewed by: Anish Nath
  • Last updated: 2025-11-19

Trust & Privacy

  • Runs locally with synthetic or provided data; nothing is uploaded.
Quick Access