/ml-model-explanation

name: ML Model Explanation
description: Interpret machine learning models using SHAP, LIME, feature importance, partial dependence, and attention visualization for explainability

ML Model Explanation

Model explainability makes machine learning decisions transparent and interpretable, enabling trust, compliance, debugging, and actionable insights from predictions.

Explanation Techniques

• **Feature Importance**: Global feature contribution to predictions
• **SHAP Values**: Game theory-based feature attribution
• **LIME**: Local linear approximations for individual predictions
• **Partial Dependence Plots**: Feature relationship with predictions
• **Attention Maps**: Visualization of model focus areas
• **Surrogate Models**: Simpler interpretable approximations

Explainability Types

• **Global**: Overall model behavior and patterns
• **Local**: Explanation for individual predictions
• **Feature-Level**: Which features matter most
• **Model-Level**: How different components interact

Python Implementation

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifier
from sklearn.linear_model import LogisticRegression
from sklearn.tree import DecisionTreeClassifier, plot_tree
from sklearn.inspection import partial_dependence, permutation_importance
import warnings
warnings.filterwarnings('ignore')

print("=== 1. Feature Importance Analysis ===")

# Create dataset
X, y = make_classification(n_samples=1000, n_features=20, n_informative=10,
                          n_redundant=5, random_state=42)
feature_names = [f'Feature_{i}' for i in range(20)]
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

# Train models
rf_model = RandomForestClassifier(n_estimators=100, random_state=42)
rf_model.fit(X_train, y_train)

gb_model = GradientBoostingClassifier(n_estimators=100, random_state=42)
gb_model.fit(X_train, y_train)

# Feature importance methods
print("\n=== Feature Importance Comparison ===")

# 1. Impurity-based importance (default)
impurity_importance = rf_model.feature_importances_

# 2. Permutation importance
perm_importance = permutation_importance(rf_model, X_test, y_test, n_repeats=10, random_state=42)

# Create comparison dataframe
importance_df = pd.DataFrame({
    'Feature': feature_names,
    'Impurity': impurity_importance,
    'Permutation': perm_importance.importances_mean
}).sort_values('Impurity', ascending=False)

print("\nTop 10 Most Important Features (by Impurity):")
print(importance_df.head(10)[['Feature', 'Impurity']])

# 2. SHAP-like Feature Attribution
print("\n=== SHAP-like Feature Attribution ===")

class SimpleShapCalculator:
    def __init__(self, model, X_background):
        self.model = model
        self.X_background = X_background
        self.baseline = model.predict_proba(X_background.mean(axis=0).reshape(1, -1))[0]

    def predict_difference(self, X_sample):
        """Get prediction difference from baseline"""
        pred = self.model.predict_proba(X_sample)[0]
        return pred - self.baseline

    def calculate_shap_values(self, X_instance, n_iterations=100):
        """Approximate SHAP values"""
        shap_values = np.zeros(X_instance.shape[1])
        n_features = X_instance.shape[1]

        for i in range(n_iterations):
            # Random feature subset
            subset_mask = np.random.random(n_features) > 0.5

            # With and without feature
            X_with = X_instance.copy()
            X_without = X_instance.copy()
            X_without[0, ~subset_mask] = self.X_background[0, ~subset_mask]

            # Marginal contribution
            contribution = (self.predict_difference(X_with)[1] -
                          self.predict_difference(X_without)[1])

            shap_values[~subset_mask] += contribution / n_iterations

        return shap_values

shap_calc = SimpleShapCalculator(rf_model, X_train)

# Calculate SHAP values for a sample
sample_idx = 0
shap_vals = shap_calc.calculate_shap_values(X_test[sample_idx:sample_idx+1], n_iterations=50)

print(f"\nSHAP Values for Sample {sample_idx}:")
shap_df = pd.DataFrame({
    'Feature': feature_names,
    'SHAP_Value': shap_vals
}).sort_values('SHAP_Value', key=abs, ascending=False)

print(shap_df.head(10)[['Feature', 'SHAP_Value']])

# 3. Partial Dependence Analysis
print("\n=== 3. Partial Dependence Analysis ===")

# Calculate partial dependence for top features
top_features = importance_df['Feature'].head(3).values
top_feature_indices = [feature_names.index(f) for f in top_features]

pd_data = {}
for feature_idx in top_feature_indices:
    pd_result = partial_dependence(rf_model, X_test, [feature_idx])
    pd_data[feature_names[feature_idx]] = pd_result

print(f"Partial dependence calculated for features: {list(pd_data.keys())}")

# 4. LIME - Local Interpretable Model-agnostic Explanations
print("\n=== 4. LIME (Local Surrogate Model) ===")

class SimpleLIME:
    def __init__(self, model, X_train):
        self.model = model
        self.X_train = X_train
        self.scaler = StandardScaler()
        self.scaler.fit(X_train)

    def explain_instance(self, instance, n_samples=1000, n_features=10):
        """Explain prediction using local linear model"""
        # Generate perturbed samples
        scaled_instance = self.scaler.transform(instance.reshape(1, -1))
        perturbations = np.random.normal(scaled_instance, 0.3, (n_samples, instance.shape[0]))

        # Get predictions
        predictions = self.model.predict_proba(perturbations)[:, 1]

        # Train local linear model
        distances = np.sum((perturbations - scaled_instance) ** 2, axis=1)
        weights = np.exp(-distances)

        # Linear regression weights
        local_mode