Skip to content

Advanced Visual Interpretation

🎯 Purpose

This guide provides advanced visual tools for analyzing the output of classification models. It complements the EDA and modeling guidebooks by focusing on post-model diagnostics, probability behavior, threshold analysis, and interpretability across binary and multiclass settings.

📉 1. Confusion Matrix (with Heatmap)

Purpose: Show accuracy and misclassification rates per class.

from sklearn.metrics import ConfusionMatrixDisplay
ConfusionMatrixDisplay.from_predictions(y_true, y_pred, cmap="Blues")

✔️ Normalize by rows to visualize per-class recall ✔️ Annotate with % or count for stakeholders

📈 2. ROC Curve (Binary / Multiclass OvR)

Purpose: Measure model's ranking ability at different thresholds.

from sklearn.metrics import roc_curve, roc_auc_score
  • Plot TPR vs FPR
  • Add AUC to legend for comparison

✔️ Use OvR format for multiclass classifiers ✔️ Use to compare multiple models on one plot

🌿 3. Precision-Recall Curve (Imbalanced Data)

Purpose: Reveal classifier performance on minority class.

from sklearn.metrics import precision_recall_curve

✔️ Steep drop-off = model sensitivity to threshold ✔️ Use area under PR curve as stability measure

📦 4. Predicted Probability Histogram

Purpose: Show model confidence across samples.

plt.hist(y_proba, bins=20)

✔️ Sharp peaks near 0 or 1 = confident classifier ⚠️ Flat or centered = underfit or poorly calibrated

📏 5. Calibration Curve (Reliability Plot)

Purpose: Compare predicted proba to actual likelihood.

from sklearn.calibration import calibration_curve

✔️ Curve ≈ diagonal = well-calibrated ⚠️ Over/underconfidence curves inform post-hoc scaling

🧪 6. Threshold Tuning Plot

Purpose: Visualize how precision, recall, and F1 change across thresholds.

from sklearn.metrics import precision_recall_curve

✔️ Identify optimal threshold (not always 0.5) ✔️ Pair with business cost matrix if available

🧠 7. SHAP / Feature Importance Plots

For Tree Models:

shap.summary_plot(shap_values, X_test)

For Linear Models:

sns.barplot(x=coefficients, y=feature_names)

✔️ Use to explain model behavior by input ✔️ Export visuals to stakeholder decks

🧭 8. Per-Class Breakdown (Multiclass Models)

Purpose: Visually diagnose class-level performance.

  • Confusion matrix heatmap (normalized)
  • One-vs-rest ROC/PR curves
  • Class-specific confidence histograms

✔️ Flag underperforming classes by color threshold ✔️ Use radar plots for class summary profiles

📋 Analyst Visual Review Checklist

  • [ ] Confusion matrix plotted and interpreted
  • [ ] ROC or PR curve plotted by class
  • [ ] Probability distribution reviewed
  • [ ] Calibration plot created
  • [ ] Threshold vs F1/Recall chart analyzed
  • [ ] SHAP or feature impact plot exported
  • [ ] Stakeholder visuals saved

💡 Final Tip

“Visuals are your interface between model truth and stakeholder understanding. Always tune thresholds and validate confidence.”

Use with: Classifier Statistical Summary Sheet, Evaluation Checklist, and Modeling Guidebook.