ML Visualizations

Visualize a confusion matrix as an annotated heatmap. Rows represent true labels, columns represent predicted labels. Cell [i,j] shows how many samples with true label i were predicted as j. Diagonal cells are correct predictions. Off-diagonal cells reveal which classes are being confused. Color intensity is proportional to count. Pass class names via labels for readable axis labels.

Plot the Receiver Operating Characteristic curve — True Positive Rate vs False Positive Rate across all classification thresholds. Includes a dashed diagonal line representing a random classifier (AUC = 0.5). If auc is provided, it is annotated on the plot. A curve hugging the top-left corner indicates strong discrimination ability. Use rocCurve() from deepbox/metrics to compute fpr and tpr.

Plot the Precision-Recall curve — Precision vs Recall across all thresholds. More informative than ROC for imbalanced datasets where the positive class is rare. A curve hugging the top-right corner indicates both high precision and high recall. If ap (average precision) is provided, it is annotated on the plot.

Plot training and validation scores as a function of training set size. The key diagnostic plot for bias-variance tradeoff. If both curves plateau at a low score → high bias (underfitting, need a more complex model). If training score is high but validation score is low → high variance (overfitting, need more data or regularization). Gap between curves = generalization gap.

Plot training and validation scores as a function of a single hyperparameter. Shows how the model behaves across different values (e.g., regularization strength C, tree depth, number of neighbors k). The optimal value is where the validation score peaks. Left of peak = underfitting, right of peak = overfitting.

Visualize the decision boundary of a 2D classifier. Creates a dense mesh grid covering the feature space, predicts the class at every grid point, and draws filled contour regions colored by predicted class. Original data points are overlaid as scatter markers. Only works with 2-feature datasets (use PCA to reduce first if needed). The model must have a .predict() method.