model-evaluation

name	model-evaluation
description	Evaluates machine learning models for performance, fairness, and reliability using appropriate metrics and validation techniques. Trigger keywords: model evaluation, metrics, accuracy, precision, recall, F1, ROC, AUC, cross-validation, ML testing. Use when this capability is needed.

Model Evaluation

Overview

This skill focuses on comprehensive evaluation of machine learning models. It covers metric selection, validation strategies, fairness assessment, and production monitoring for ensuring model quality and reliability.

Instructions

1. Define Evaluation Criteria

• Identify business objectives
• Select appropriate metrics
• Define success thresholds
• Consider fairness requirements

2. Design Evaluation Strategy

• Choose validation approach
• Plan for data splits
• Handle class imbalance
• Account for temporal aspects

3. Conduct Evaluation

• Calculate performance metrics
• Analyze error patterns
• Assess model fairness
• Test edge cases

4. Report and Monitor

• Document evaluation results
• Create monitoring dashboards
• Set up alerting thresholds
• Plan for retraining

Best Practices

1. Match Metrics to Goals: Choose metrics aligned with business objectives
2. Use Multiple Metrics: No single metric tells the whole story
3. Proper Validation: Use appropriate cross-validation schemes
4. Test Distribution Shift: Evaluate on out-of-distribution data
5. Check for Bias: Assess fairness across demographic groups
6. Version Everything: Track models, data, and metrics
7. Monitor Production: Continuously track model performance

Examples

Example 1: Classification Model Evaluation

import numpy as np
import pandas as pd
from sklearn.metrics import (
    accuracy_score, precision_score, recall_score, f1_score,
    roc_auc_score, average_precision_score, confusion_matrix,
    classification_report, roc_curve, precision_recall_curve
)
import matplotlib.pyplot as plt

class ClassificationEvaluator:
    """Comprehensive classification model evaluator."""

    def __init__(self, y_true, y_pred, y_prob=None, class_names=None):
        self.y_true = y_true
        self.y_pred = y_pred
        self.y_prob = y_prob
        self.class_names = class_names or ['Negative', 'Positive']

    def compute_metrics(self) -> dict:
        """Compute all classification metrics."""
        metrics = {
            'accuracy': accuracy_score(self.y_true, self.y_pred),
            'precision': precision_score(self.y_true, self.y_pred, average='weighted'),
            'recall': recall_score(self.y_true, self.y_pred, average='weighted'),
            'f1': f1_score(self.y_true, self.y_pred, average=),
        }

         .y_prob   :
            metrics[] = roc_auc_score(.y_true, .y_prob)
            metrics[] = average_precision_score(.y_true, .y_prob)

         metrics

     () -> :
        
        cm = confusion_matrix(.y_true, .y_pred)
        tn, fp, fn, tp = cm.ravel()

         {
            : cm,
            : tn,
            : fp,
            : fn,
            : tp,
            : tn / (tn + fp),
            : tp / (tp + fn),
            : fp / (fp + tn),
            : fn / (fn + tp),
        }

     ():
        
         .y_prob  :
             ValueError()

        fpr, tpr, thresholds = roc_curve(.y_true, .y_prob)
        auc = roc_auc_score(.y_true, .y_prob)

        plt.figure(figsize=(, ))
        plt.plot(fpr, tpr, label=)
        plt.plot([, ], [, ], , label=)
        plt.xlabel()
        plt.ylabel()
        plt.title()
        plt.legend()
        plt.grid(, alpha=)

         save_path:
            plt.savefig(save_path, dpi=, bbox_inches=)
        plt.show()

     () -> :
        
        metrics = .compute_metrics()
        cm_analysis = .confusion_matrix_analysis()

        report = 
         report


evaluator = ClassificationEvaluator(y_true, y_pred, y_prob)
(evaluator.generate_report())
evaluator.plot_roc_curve()

Example 2: Regression Model Evaluation

from sklearn.metrics import (
    mean_squared_error, mean_absolute_error, r2_score,
    mean_absolute_percentage_error, explained_variance_score
)
import numpy as np

class RegressionEvaluator:
    """Comprehensive regression model evaluator."""

    def __init__(self, y_true, y_pred):
        self.y_true = np.array(y_true)
        self.y_pred = np.array(y_pred)
        self.residuals = self.y_true - self.y_pred

    def compute_metrics(self) -> dict:
        """Compute all regression metrics."""
        mse = mean_squared_error(self.y_true, self.y_pred)

        return {
            'mse': mse,
            'rmse': np.sqrt(mse),
            'mae': mean_absolute_error(self.y_true, self.y_pred),
            'mape': mean_absolute_percentage_error(self.y_true, self.y_pred) * 100,
            'r2': r2_score(self.y_true, self.y_pred),
            'explained_variance': explained_variance_score(self.y_true, self.y_pred),
        }

    def residual_analysis(self) -> dict:
        
         {
            : np.mean(.residuals),
            : np.std(.residuals),
            : np.(.residuals),
            : np.(.residuals),
            : ._skewness(.residuals),
        }

     ():
        
        n = (data)
        mean = np.mean(data)
        std = np.std(data)
         (n / ((n-) * (n-))) * np.(((data - mean) / std) ** )

     ():
        
        fig, axes = plt.subplots(, , figsize=(, ))

        
        ax1 = axes[, ]
        ax1.scatter(.y_true, .y_pred, alpha=)
        ax1.plot([.y_true.(), .y_true.()],
                 [.y_true.(), .y_true.()], )
        ax1.set_xlabel()
        ax1.set_ylabel()
        ax1.set_title()

        
        ax2 = axes[, ]
        ax2.scatter(.y_pred, .residuals, alpha=)
        ax2.axhline(y=, color=, linestyle=)
        ax2.set_xlabel()
        ax2.set_ylabel()
        ax2.set_title()

        
        ax3 = axes[, ]
        ax3.hist(.residuals, bins=, edgecolor=)
        ax3.set_xlabel()
        ax3.set_ylabel()
        ax3.set_title()

        
        ax4 = axes[, ]
         scipy  stats
        stats.probplot(.residuals, dist=, plot=ax4)
        ax4.set_title()

        plt.tight_layout()
         save_path:
            plt.savefig(save_path, dpi=, bbox_inches=)
        plt.show()

Example 3: Cross-Validation Strategies

from sklearn.model_selection import (
    cross_val_score, StratifiedKFold, TimeSeriesSplit,
    GroupKFold, cross_validate
)

def evaluate_with_cv(model, X, y, cv_strategy='stratified', n_splits=5, groups=None):
    """
    Evaluate model with appropriate cross-validation strategy.

    Args:
        model: Sklearn-compatible model
        X: Features
        y: Target
        cv_strategy: 'stratified', 'timeseries', 'group', or 'kfold'
        n_splits: Number of CV folds
        groups: Group labels for GroupKFold

    Returns:
        Dictionary with CV results
    """
    # Select CV strategy
    if cv_strategy == 'stratified':
        cv = StratifiedKFold(n_splits=n_splits, shuffle=True, random_state=42)
    elif cv_strategy == 'timeseries':
        cv = TimeSeriesSplit(n_splits=n_splits)
    elif cv_strategy == 'group':
        cv = GroupKFold(n_splits=n_splits)
    else:
        cv = n_splits

    # Define scoring metrics
    scoring = {
        'accuracy': 'accuracy',
        'precision': 'precision_weighted',
        'recall': 'recall_weighted',
        'f1': 'f1_weighted',
        'roc_auc': 'roc_auc'
    }

    # Perform cross-validation
    cv_results = cross_validate(
        model, X, y,
        cv=cv,
        scoring=scoring,
        groups=groups,
        return_train_score=True,
        n_jobs=-1
    )

    # Summarize results
    summary = {}
    for metric in scoring.keys():
        test_scores = cv_results[]
        train_scores = cv_results[]
        summary[metric] = {
            : np.mean(test_scores),
            : np.std(test_scores),
            : np.mean(train_scores),
            : np.std(train_scores),
            : np.mean(train_scores) - np.mean(test_scores)
        }

     summary


results = evaluate_with_cv(model, X, y, cv_strategy=, n_splits=)
 metric, values  results.items():
    ()

Example 4: Fairness Evaluation

def evaluate_fairness(y_true, y_pred, sensitive_attr, favorable_label=1):
    """
    Evaluate model fairness across demographic groups.

    Args:
        y_true: True labels
        y_pred: Predicted labels
        sensitive_attr: Protected attribute values
        favorable_label: The favorable outcome label

    Returns:
        Dictionary with fairness metrics
    """
    groups = np.unique(sensitive_attr)
    results = {'group_metrics': {}}

    for group in groups:
        mask = sensitive_attr == group
        group_true = y_true[mask]
        group_pred = y_pred[mask]

        # Calculate group-specific metrics
        tp = np.sum((group_true == favorable_label) & (group_pred == favorable_label))
        fp = np.sum((group_true != favorable_label) & (group_pred == favorable_label))
        fn = np.sum((group_true == favorable_label) & (group_pred != favorable_label))
        tn = np.sum((group_true != favorable_label) & (group_pred != favorable_label))

        results['group_metrics'][group] = {
            'selection_rate': np.mean(group_pred == favorable_label),
            'tpr': tp / (tp + fn) if (tp + fn) > 0 else 0,
            'fpr': fp / (fp + tn) if (fp + tn) > 0 else 0,
            'accuracy': np.mean(group_true == group_pred),
            'size': len(group_true)
        }

    # Calculate fairness metrics
    selection_rates = [m['selection_rate'] for m in results['group_metrics'].values()]
    tprs = [m['tpr'] for m in results[].values()]
    fprs = [m[]  m  results[].values()]

    results[] = {
        : (selection_rates) - (selection_rates),
        : (tprs) - (tprs),
        : (fprs) - (fprs),
    }

     results

---

Source: ForceInjection/domain-driven-design-skills — distributed by TomeVault.