explainiverse 0.8.1__py3-none-any.whl → 0.8.2__py3-none-any.whl

explainiverse/__init__.py

@@ -34,7 +34,7 @@ from explainiverse.adapters.sklearn_adapter import SklearnAdapter
 from explainiverse.adapters import TORCH_AVAILABLE
 from explainiverse.engine.suite import ExplanationSuite
 
-__version__ = "0.8.1"
+__version__ = "0.8.2"
 
 __all__ = [
     # Core

explainiverse/evaluation/__init__.py

@@ -3,9 +3,10 @@
 Evaluation metrics for explanation quality.
 
 Includes:
-- Faithfulness metrics (PGI, PGU, Comprehensiveness, Sufficiency)
+- Faithfulness metrics (PGI, PGU, Comprehensiveness, Sufficiency, Faithfulness Estimate)
 - Stability metrics (RIS, ROS, Lipschitz)
 - Perturbation metrics (AOPC, ROAR)
+- Extended faithfulness metrics (Phase 1 expansion)
 """
 
 from explainiverse.evaluation.metrics import (
@@ -35,13 +36,20 @@ from explainiverse.evaluation.stability import (
     compare_explainer_stability,
 )
 
+from explainiverse.evaluation.faithfulness_extended import (
+    compute_faithfulness_estimate,
+    compute_batch_faithfulness_estimate,
+    compute_monotonicity,
+    compute_batch_monotonicity,
+)
+
 __all__ = [
     # Perturbation metrics (existing)
     "compute_aopc",
     "compute_batch_aopc",
     "compute_roar",
     "compute_roar_curve",
-    # Faithfulness metrics (new)
+    # Faithfulness metrics (core)
     "compute_pgi",
     "compute_pgu",
     "compute_faithfulness_score",
@@ -50,11 +58,16 @@ __all__ = [
     "compute_faithfulness_correlation",
     "compare_explainer_faithfulness",
     "compute_batch_faithfulness",
-    # Stability metrics (new)
+    # Stability metrics
     "compute_ris",
     "compute_ros",
     "compute_lipschitz_estimate",
     "compute_stability_metrics",
     "compute_batch_stability",
     "compare_explainer_stability",
+    # Extended faithfulness metrics (Phase 1)
+    "compute_faithfulness_estimate",
+    "compute_batch_faithfulness_estimate",
+    "compute_monotonicity",
+    "compute_batch_monotonicity",
 ]

explainiverse/evaluation/faithfulness_extended.py

@@ -0,0 +1,416 @@
+# src/explainiverse/evaluation/faithfulness_extended.py
+"""
+Extended faithfulness evaluation metrics.
+
+Phase 1 metrics for exceeding OpenXAI/Quantus:
+- Faithfulness Estimate (Alvarez-Melis et al., 2018)
+- Monotonicity (Arya et al., 2019)
+- Monotonicity-Nguyen (Nguyen et al., 2020)
+- Pixel Flipping (Bach et al., 2015)
+- Region Perturbation (Samek et al., 2015)
+- Selectivity (Montavon et al., 2018)
+- Sensitivity-n (Ancona et al., 2018)
+- IROF (Rieger & Hansen, 2020)
+- Infidelity (Yeh et al., 2019)
+- ROAD (Rong et al., 2022)
+- Insertion AUC (Petsiuk et al., 2018)
+- Deletion AUC (Petsiuk et al., 2018)
+"""
+import numpy as np
+import re
+from typing import Union, Callable, List, Dict, Optional, Tuple
+from scipy import stats
+
+from explainiverse.core.explanation import Explanation
+from explainiverse.evaluation._utils import (
+    get_sorted_feature_indices,
+    compute_baseline_values,
+    apply_feature_mask,
+    resolve_k,
+    get_prediction_value,
+    compute_prediction_change,
+)
+
+
+def _extract_attribution_array(
+    explanation: Explanation,
+    n_features: int
+) -> np.ndarray:
+    """
+    Extract attribution values as a numpy array in feature index order.
+    
+    Args:
+        explanation: Explanation object with feature_attributions
+        n_features: Expected number of features
+        
+    Returns:
+        1D numpy array of attribution values ordered by feature index
+    """
+    attributions = explanation.explanation_data.get("feature_attributions", {})
+    feature_names = getattr(explanation, 'feature_names', None)
+    
+    if not attributions:
+        raise ValueError("No feature attributions found in explanation.")
+    
+    # Build attribution array in feature order
+    attr_array = np.zeros(n_features)
+    
+    if feature_names is not None:
+        for fname, value in attributions.items():
+            # Try to find the index for this feature name
+            for i, fn in enumerate(feature_names):
+                if fn == fname or fn in fname or fname in fn:
+                    attr_array[i] = value
+                    break
+            else:
+                # Try extracting index from name pattern
+                for pattern in [r'feature[_\s]*(\d+)', r'feat[_\s]*(\d+)', r'^f(\d+)', r'^x(\d+)']:
+                    match = re.search(pattern, fname, re.IGNORECASE)
+                    if match:
+                        idx = int(match.group(1))
+                        if 0 <= idx < n_features:
+                            attr_array[idx] = value
+                        break
+    else:
+        # No feature names - try to extract indices from keys
+        for fname, value in attributions.items():
+            for pattern in [r'feature[_\s]*(\d+)', r'feat[_\s]*(\d+)', r'^f(\d+)', r'^x(\d+)']:
+                match = re.search(pattern, fname, re.IGNORECASE)
+                if match:
+                    idx = int(match.group(1))
+                    if 0 <= idx < n_features:
+                        attr_array[idx] = value
+                    break
+    
+    return attr_array
+
+
+# =============================================================================
+# Metric 1: Faithfulness Estimate (Alvarez-Melis & Jaakkola, 2018)
+# =============================================================================
+
+def compute_faithfulness_estimate(
+    model,
+    instance: np.ndarray,
+    explanation: Explanation,
+    baseline: Union[str, float, np.ndarray, Callable] = "mean",
+    background_data: np.ndarray = None,
+    subset_size: int = None,
+    n_subsets: int = 100,
+    seed: int = None,
+) -> float:
+    """
+    Compute Faithfulness Estimate (Alvarez-Melis & Jaakkola, 2018).
+    
+    Measures the correlation between feature attributions and the actual
+    impact on predictions when individual features are perturbed. For each
+    feature, computes the prediction change when that feature is replaced
+    with baseline, then correlates these changes with attribution magnitudes.
+    
+    Higher correlation indicates the explanation correctly identifies
+    which features actually matter for the prediction.
+    
+    Args:
+        model: Model adapter with predict/predict_proba method
+        instance: Input instance (1D array)
+        explanation: Explanation object with feature_attributions
+        baseline: Baseline for feature replacement ("mean", "median", scalar, array, callable)
+        background_data: Reference data for computing baseline (required for "mean"/"median")
+        subset_size: Size of random subsets to perturb (default: 1 for single-feature)
+        n_subsets: Number of random subsets to evaluate (used when subset_size > 1)
+        seed: Random seed for reproducibility
+        
+    Returns:
+        Faithfulness estimate score (Pearson correlation, -1 to 1, higher is better)
+        
+    References:
+        Alvarez-Melis, D., & Jaakkola, T. S. (2018). Towards Robust Interpretability 
+        with Self-Explaining Neural Networks. NeurIPS.
+    """
+    if seed is not None:
+        np.random.seed(seed)
+    
+    instance = np.asarray(instance).flatten()
+    n_features = len(instance)
+    
+    # Get baseline values
+    baseline_values = compute_baseline_values(
+        baseline, background_data, n_features
+    )
+    
+    # Extract attributions as array
+    attr_array = _extract_attribution_array(explanation, n_features)
+    
+    # Default subset_size is 1 (single-feature perturbation)
+    if subset_size is None:
+        subset_size = 1
+    
+    if subset_size == 1:
+        # Single-feature perturbation: evaluate each feature individually
+        prediction_changes = []
+        attribution_values = []
+        
+        for i in range(n_features):
+            # Skip features with zero attribution (they won't affect correlation)
+            if abs(attr_array[i]) < 1e-10:
+                continue
+            
+            # Perturb single feature
+            perturbed = apply_feature_mask(instance, [i], baseline_values)
+            
+            # Compute prediction change
+            change = compute_prediction_change(model, instance, perturbed, metric="absolute")
+            
+            prediction_changes.append(change)
+            attribution_values.append(abs(attr_array[i]))
+        
+        if len(prediction_changes) < 2:
+            return 0.0  # Not enough data points for correlation
+        
+        # Compute Pearson correlation
+        corr, _ = stats.pearsonr(attribution_values, prediction_changes)
+        
+        return float(corr) if not np.isnan(corr) else 0.0
+    
+    else:
+        # Random subset perturbation
+        prediction_changes = []
+        attribution_sums = []
+        
+        for _ in range(n_subsets):
+            # Sample random subset of features
+            subset_indices = np.random.choice(
+                n_features, size=min(subset_size, n_features), replace=False
+            )
+            
+            # Perturb subset
+            perturbed = apply_feature_mask(instance, subset_indices.tolist(), baseline_values)
+            
+            # Compute prediction change
+            change = compute_prediction_change(model, instance, perturbed, metric="absolute")
+            
+            # Sum of attributions in subset
+            attr_sum = np.sum(np.abs(attr_array[subset_indices]))
+            
+            prediction_changes.append(change)
+            attribution_sums.append(attr_sum)
+        
+        if len(prediction_changes) < 2:
+            return 0.0
+        
+        # Compute Pearson correlation
+        corr, _ = stats.pearsonr(attribution_sums, prediction_changes)
+        
+        return float(corr) if not np.isnan(corr) else 0.0
+
+
+def compute_batch_faithfulness_estimate(
+    model,
+    X: np.ndarray,
+    explanations: List[Explanation],
+    baseline: Union[str, float, np.ndarray, Callable] = "mean",
+    max_samples: int = None,
+    seed: int = None,
+) -> Dict[str, float]:
+    """
+    Compute average Faithfulness Estimate over a batch of instances.
+    
+    Args:
+        model: Model adapter
+        X: Input data (2D array)
+        explanations: List of Explanation objects (one per instance)
+        baseline: Baseline for feature replacement
+        max_samples: Maximum number of samples to evaluate
+        seed: Random seed
+        
+    Returns:
+        Dictionary with mean, std, min, max, and count of valid scores
+    """
+    n_samples = len(explanations)
+    if max_samples:
+        n_samples = min(n_samples, max_samples)
+    
+    scores = []
+    
+    for i in range(n_samples):
+        try:
+            score = compute_faithfulness_estimate(
+                model, X[i], explanations[i],
+                baseline=baseline, background_data=X,
+                seed=seed
+            )
+            if not np.isnan(score):
+                scores.append(score)
+        except Exception:
+            continue
+    
+    if not scores:
+        return {"mean": 0.0, "std": 0.0, "min": 0.0, "max": 0.0, "n_samples": 0}
+    
+    return {
+        "mean": float(np.mean(scores)),
+        "std": float(np.std(scores)),
+        "min": float(np.min(scores)),
+        "max": float(np.max(scores)),
+        "n_samples": len(scores),
+    }
+
+
+# =============================================================================
+# Metric 2: Monotonicity (Arya et al., 2019)
+# =============================================================================
+
+def compute_monotonicity(
+    model,
+    instance: np.ndarray,
+    explanation: Explanation,
+    baseline: Union[str, float, np.ndarray, Callable] = "mean",
+    background_data: np.ndarray = None,
+    target_class: int = None,
+    use_absolute: bool = True,
+    tolerance: float = 1e-6,
+) -> float:
+    """
+    Compute Monotonicity (Arya et al., 2019).
+    
+    Measures whether sequentially adding features in order of their attributed
+    importance monotonically increases the model's prediction confidence.
+    Starting from a baseline (all features masked), features are revealed
+    one-by-one in descending order of attribution. A faithful explanation
+    should show monotonically increasing predictions.
+    
+    Args:
+        model: Model adapter with predict/predict_proba method
+        instance: Input instance (1D array)
+        explanation: Explanation object with feature_attributions
+        baseline: Baseline for masked features ("mean", "median", scalar, array, callable)
+        background_data: Reference data for computing baseline (required for "mean"/"median")
+        target_class: Target class index for probability (default: predicted class)
+        use_absolute: If True, sort features by absolute attribution value
+        tolerance: Small value for numerical stability in monotonicity check
+        
+    Returns:
+        Monotonicity score (0 to 1, higher is better)
+        1.0 means perfectly monotonic increase
+        
+    References:
+        Arya, V., et al. (2019). One Explanation Does Not Fit All: A Toolkit and 
+        Taxonomy of AI Explainability Techniques. arXiv:1909.03012.
+    """
+    instance = np.asarray(instance).flatten()
+    n_features = len(instance)
+    
+    # Get baseline values
+    baseline_values = compute_baseline_values(
+        baseline, background_data, n_features
+    )
+    
+    # Extract attributions as array
+    attr_array = _extract_attribution_array(explanation, n_features)
+    
+    # Sort features by attribution (descending - most important first)
+    if use_absolute:
+        sorted_indices = np.argsort(-np.abs(attr_array))
+    else:
+        sorted_indices = np.argsort(-attr_array)
+    
+    # Determine target class
+    if target_class is None:
+        # Use predicted class
+        pred = get_prediction_value(model, instance.reshape(1, -1))
+        if isinstance(pred, np.ndarray) and pred.ndim > 0:
+            target_class = int(np.argmax(pred))
+        else:
+            target_class = 0
+    
+    # Start from baseline (all features masked)
+    current = baseline_values.copy()
+    
+    # Track predictions as features are revealed
+    predictions = []
+    
+    # Get initial prediction (baseline state)
+    pred = get_prediction_value(model, current.reshape(1, -1))
+    if isinstance(pred, np.ndarray) and pred.ndim > 0 and len(pred) > target_class:
+        predictions.append(pred[target_class])
+    else:
+        predictions.append(float(pred))
+    
+    # Add features one by one
+    revealed_features = []
+    for idx in sorted_indices:
+        # Reveal this feature (set to original value)
+        revealed_features.append(idx)
+        current[idx] = instance[idx]
+        
+        # Get prediction
+        pred = get_prediction_value(model, current.reshape(1, -1))
+        if isinstance(pred, np.ndarray) and pred.ndim > 0 and len(pred) > target_class:
+            predictions.append(pred[target_class])
+        else:
+            predictions.append(float(pred))
+    
+    # Count monotonic increases
+    # A step is monotonic if: pred[i+1] >= pred[i] - tolerance
+    n_steps = len(predictions) - 1
+    if n_steps == 0:
+        return 1.0
+    
+    monotonic_steps = 0
+    for i in range(n_steps):
+        if predictions[i + 1] >= predictions[i] - tolerance:
+            monotonic_steps += 1
+    
+    return float(monotonic_steps) / float(n_steps)
+
+
+def compute_batch_monotonicity(
+    model,
+    X: np.ndarray,
+    explanations: List[Explanation],
+    baseline: Union[str, float, np.ndarray, Callable] = "mean",
+    max_samples: int = None,
+    use_absolute: bool = True,
+) -> Dict[str, float]:
+    """
+    Compute average Monotonicity over a batch of instances.
+    
+    Args:
+        model: Model adapter
+        X: Input data (2D array)
+        explanations: List of Explanation objects (one per instance)
+        baseline: Baseline for masked features
+        max_samples: Maximum number of samples to evaluate
+        use_absolute: If True, sort features by absolute attribution value
+        
+    Returns:
+        Dictionary with mean, std, min, max, and count of valid scores
+    """
+    n_samples = len(explanations)
+    if max_samples:
+        n_samples = min(n_samples, max_samples)
+    
+    scores = []
+    
+    for i in range(n_samples):
+        try:
+            score = compute_monotonicity(
+                model, X[i], explanations[i],
+                baseline=baseline, background_data=X,
+                use_absolute=use_absolute
+            )
+            if not np.isnan(score):
+                scores.append(score)
+        except Exception:
+            continue
+    
+    if not scores:
+        return {"mean": 0.0, "std": 0.0, "min": 0.0, "max": 0.0, "n_samples": 0}
+    
+    return {
+        "mean": float(np.mean(scores)),
+        "std": float(np.std(scores)),
+        "min": float(np.min(scores)),
+        "max": float(np.max(scores)),
+        "n_samples": len(scores),
+    }

{explainiverse-0.8.1.dist-info → explainiverse-0.8.2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: explainiverse
-Version: 0.8.1
+Version: 0.8.2
 Summary: Unified, extensible explainability framework supporting 18 XAI methods including LIME, SHAP, LRP, TCAV, GradCAM, and more
 Home-page: https://github.com/jemsbhai/explainiverse
 License: MIT

{explainiverse-0.8.1.dist-info → explainiverse-0.8.2.dist-info}/RECORD

@@ -1,4 +1,4 @@
-explainiverse/__init__.py,sha256=fnB58cF7z1PvBjNkKvEcmXI0u2l3T3FYLyCx1sLEmXQ,1694
+explainiverse/__init__.py,sha256=icvNmaSq0DAERqIrU60N60KCIspHbtEWTi3kt_YXTUI,1694
 explainiverse/adapters/__init__.py,sha256=HcQGISyp-YQ4jEj2IYveX_c9X5otLcTNWRnVRRhzRik,781
 explainiverse/adapters/base_adapter.py,sha256=Nqt0GeDn_-PjTyJcZsE8dRTulavqFQsv8sMYWS_ps-M,603
 explainiverse/adapters/pytorch_adapter.py,sha256=DLQKJ7gB0foPwAmcrru7QdZnPRnhqDKpFCT-EaD3420,15612
@@ -9,9 +9,10 @@ explainiverse/core/explanation.py,sha256=498BbRYrNR-BOql78sENOsyWxgqLsBVZXn14lh-
 explainiverse/core/registry.py,sha256=6HttL27Ty4jYtugRf-EDIKPy80M8BfvUppAKwwGDyQ8,27207
 explainiverse/engine/__init__.py,sha256=1sZO8nH1mmwK2e-KUavBQm7zYDWUe27nyWoFy9tgsiA,197
 explainiverse/engine/suite.py,sha256=G-7OjESisSTaQ1FQrlPl4YydX13uz8Bb70hJZNlcl2M,8918
-explainiverse/evaluation/__init__.py,sha256=ePE97KwSjg_IChZ03DeQax8GruTjx-BVrMSi_nzoyoA,1501
+explainiverse/evaluation/__init__.py,sha256=XFVnmwrRtHHhtxI_yOw_nsR67pJvH-IBO_lEUVI-eDE,1957
 explainiverse/evaluation/_utils.py,sha256=ej7YOPZ90gVHuuIMj45EXHq9Jx3QG7lhaj5sk26hRpg,10519
 explainiverse/evaluation/faithfulness.py,sha256=_40afOW6vJ3dQguHlJySlgWqiJF_xIvN-uVA3nPKRvI,14841
+explainiverse/evaluation/faithfulness_extended.py,sha256=0zHcmINNA88EJcKOY04Z384S3QhBMo7W2m3lGNkUiNQ,14690
 explainiverse/evaluation/metrics.py,sha256=snNK9Ua1VzHDT6DlrhYL4m2MmRF3X15vuuVXiHbeicU,9944
 explainiverse/evaluation/stability.py,sha256=q2d3rpxpp0X1s6ADST1iZA4tzksLJpR0mYBnA_U5FIs,12090
 explainiverse/explainers/__init__.py,sha256=-ncRXbFKahH3bR0oXM2UQM4LtTdTlvdeprL6cHeqNBs,2549
@@ -38,7 +39,7 @@ explainiverse/explainers/gradient/smoothgrad.py,sha256=COIKZSFcApmMkA62M0AForHiY
 explainiverse/explainers/gradient/tcav.py,sha256=zc-8wMsc2ZOhUeSZNBJ6H6BPXlVMJ9DRcAMiL25wU9I,32242
 explainiverse/explainers/rule_based/__init__.py,sha256=gKzlFCAzwurAMLJcuYgal4XhDj1thteBGcaHWmN7iWk,243
 explainiverse/explainers/rule_based/anchors_wrapper.py,sha256=ML7W6aam-eMGZHy5ilol8qupZvNBJpYAFatEEPnuMyo,13254
-explainiverse-0.8.1.dist-info/LICENSE,sha256=28rbHe8rJgmUlRdxJACfq1Sj-MtCEhyHxkJedQd1ZYA,1070
-explainiverse-0.8.1.dist-info/METADATA,sha256=A5NjgO7v2I-P8sjFW8yaJCd_VcU_Dqumq2iRdl0_Fbc,23770
-explainiverse-0.8.1.dist-info/WHEEL,sha256=sP946D7jFCHeNz5Iq4fL4Lu-PrWrFsgfLXbbkciIZwg,88
-explainiverse-0.8.1.dist-info/RECORD,,
+explainiverse-0.8.2.dist-info/LICENSE,sha256=28rbHe8rJgmUlRdxJACfq1Sj-MtCEhyHxkJedQd1ZYA,1070
+explainiverse-0.8.2.dist-info/METADATA,sha256=QSLwIr4RmoHpxqIfoarJX17alA-0esXfdNa1cemWu5s,23770
+explainiverse-0.8.2.dist-info/WHEEL,sha256=sP946D7jFCHeNz5Iq4fL4Lu-PrWrFsgfLXbbkciIZwg,88
+explainiverse-0.8.2.dist-info/RECORD,,