explainiverse 0.2.0__tar.gz → 0.2.1__tar.gz

{explainiverse-0.2.0 → explainiverse-0.2.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: explainiverse
-Version: 0.2.0
+Version: 0.2.1
 Summary: Unified, extensible explainability framework supporting LIME, SHAP, Anchors, Counterfactuals, PDP, ALE, SAGE, and more
 Home-page: https://github.com/jemsbhai/explainiverse
 License: MIT

{explainiverse-0.2.0 → explainiverse-0.2.1}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "explainiverse"
-version = "0.2.0"
+version = "0.2.1"
 description = "Unified, extensible explainability framework supporting LIME, SHAP, Anchors, Counterfactuals, PDP, ALE, SAGE, and more"
 authors = ["Muntaser Syed <jemsbhai@gmail.com>"]
 license = "MIT"

{explainiverse-0.2.0 → explainiverse-0.2.1}/src/explainiverse/__init__.py

@@ -27,7 +27,7 @@ from explainiverse.core.registry import (
 from explainiverse.adapters.sklearn_adapter import SklearnAdapter
 from explainiverse.engine.suite import ExplanationSuite
 
-__version__ = "0.2.0"
+__version__ = "0.2.1"
 
 __all__ = [
     # Core

{explainiverse-0.2.0 → explainiverse-0.2.1}/src/explainiverse/core/registry.py

@@ -362,6 +362,7 @@ def _create_default_registry() -> ExplainerRegistry:
     """Create and populate the default global registry."""
     from explainiverse.explainers.attribution.lime_wrapper import LimeExplainer
     from explainiverse.explainers.attribution.shap_wrapper import ShapExplainer
+    from explainiverse.explainers.attribution.treeshap_wrapper import TreeShapExplainer
     from explainiverse.explainers.rule_based.anchors_wrapper import AnchorsExplainer
     from explainiverse.explainers.global_explainers.permutation_importance import PermutationImportanceExplainer
     from explainiverse.explainers.global_explainers.partial_dependence import PartialDependenceExplainer
@@ -409,6 +410,23 @@ def _create_default_registry() -> ExplainerRegistry:
         )
     )
     
+    # Register TreeSHAP (optimized for tree models)
+    registry.register(
+        name="treeshap",
+        explainer_class=TreeShapExplainer,
+        meta=ExplainerMeta(
+            scope="local",
+            model_types=["tree", "ensemble"],
+            data_types=["tabular"],
+            task_types=["classification", "regression"],
+            description="TreeSHAP - exact SHAP values for tree-based models (RandomForest, XGBoost, etc.)",
+            paper_reference="Lundberg et al., 2018 - 'Consistent Individualized Feature Attribution for Tree Ensembles'",
+            complexity="O(TLD^2) - polynomial in tree depth",
+            requires_training_data=False,
+            supports_batching=True
+        )
+    )
+    
     # Register Anchors
     registry.register(
         name="anchors",

{explainiverse-0.2.0 → explainiverse-0.2.1}/src/explainiverse/explainers/__init__.py

@@ -4,7 +4,8 @@ Explainiverse Explainers - comprehensive XAI method implementations.
 
 Local Explainers (instance-level):
 - LIME: Local Interpretable Model-agnostic Explanations
-- SHAP: SHapley Additive exPlanations
+- SHAP: SHapley Additive exPlanations (KernelSHAP - model-agnostic)
+- TreeSHAP: Optimized exact SHAP for tree-based models
 - Anchors: High-precision rule-based explanations
 - Counterfactual: Diverse counterfactual explanations
 
@@ -17,6 +18,7 @@ Global Explainers (model-level):
 
 from explainiverse.explainers.attribution.lime_wrapper import LimeExplainer
 from explainiverse.explainers.attribution.shap_wrapper import ShapExplainer
+from explainiverse.explainers.attribution.treeshap_wrapper import TreeShapExplainer
 from explainiverse.explainers.rule_based.anchors_wrapper import AnchorsExplainer
 from explainiverse.explainers.counterfactual.dice_wrapper import CounterfactualExplainer
 from explainiverse.explainers.global_explainers.permutation_importance import PermutationImportanceExplainer
@@ -28,6 +30,7 @@ __all__ = [
     # Local explainers
     "LimeExplainer",
     "ShapExplainer",
+    "TreeShapExplainer",
     "AnchorsExplainer",
     "CounterfactualExplainer",
     # Global explainers

{explainiverse-0.2.0 → explainiverse-0.2.1}/src/explainiverse/explainers/attribution/__init__.py

@@ -5,5 +5,6 @@ Attribution-based explainers - feature importance explanations.
 
 from explainiverse.explainers.attribution.lime_wrapper import LimeExplainer
 from explainiverse.explainers.attribution.shap_wrapper import ShapExplainer
+from explainiverse.explainers.attribution.treeshap_wrapper import TreeShapExplainer
 
-__all__ = ["LimeExplainer", "ShapExplainer"]
+__all__ = ["LimeExplainer", "ShapExplainer", "TreeShapExplainer"]

explainiverse-0.2.1/src/explainiverse/explainers/attribution/treeshap_wrapper.py

@@ -0,0 +1,434 @@
+# src/explainiverse/explainers/attribution/treeshap_wrapper.py
+"""
+TreeSHAP Explainer - Optimized SHAP for Tree-based Models.
+
+TreeSHAP computes exact SHAP values in polynomial time for tree-based models,
+making it significantly faster than KernelSHAP while providing exact (not
+approximate) Shapley values.
+
+Reference:
+    Lundberg, S.M., Erion, G.G., & Lee, S.I. (2018). Consistent Individualized
+    Feature Attribution for Tree Ensembles. arXiv:1802.03888.
+    
+Supported Models:
+    - scikit-learn: RandomForest, GradientBoosting, DecisionTree, ExtraTrees
+    - XGBoost: XGBClassifier, XGBRegressor
+    - LightGBM: LGBMClassifier, LGBMRegressor (if installed)
+    - CatBoost: CatBoostClassifier, CatBoostRegressor (if installed)
+"""
+
+import numpy as np
+import shap
+from typing import List, Optional, Union
+
+from explainiverse.core.explainer import BaseExplainer
+from explainiverse.core.explanation import Explanation
+
+
+# Tree-based model types that TreeSHAP supports
+SUPPORTED_TREE_MODELS = (
+    "RandomForestClassifier",
+    "RandomForestRegressor",
+    "GradientBoostingClassifier",
+    "GradientBoostingRegressor",
+    "DecisionTreeClassifier",
+    "DecisionTreeRegressor",
+    "ExtraTreesClassifier",
+    "ExtraTreesRegressor",
+    "XGBClassifier",
+    "XGBRegressor",
+    "XGBRFClassifier",
+    "XGBRFRegressor",
+    "LGBMClassifier",
+    "LGBMRegressor",
+    "CatBoostClassifier",
+    "CatBoostRegressor",
+    "HistGradientBoostingClassifier",
+    "HistGradientBoostingRegressor",
+)
+
+
+def _is_tree_model(model) -> bool:
+    """Check if a model is a supported tree-based model."""
+    model_name = type(model).__name__
+    return model_name in SUPPORTED_TREE_MODELS
+
+
+def _get_raw_model(model):
+    """
+    Extract the raw model from an adapter if necessary.
+    
+    TreeExplainer needs the actual sklearn/xgboost model, not an adapter.
+    """
+    # If it's an adapter, get the underlying model
+    if hasattr(model, 'model'):
+        return model.model
+    return model
+
+
+class TreeShapExplainer(BaseExplainer):
+    """
+    TreeSHAP explainer for tree-based models.
+    
+    Uses SHAP's TreeExplainer to compute exact SHAP values in polynomial time.
+    This is significantly faster than KernelSHAP for supported tree models
+    and provides exact Shapley values rather than approximations.
+    
+    Key advantages over KernelSHAP:
+    - Exact SHAP values (not approximations)
+    - O(TLD²) complexity vs O(TL2^M) for KernelSHAP
+    - Can compute interaction values
+    - No background data sampling needed
+    
+    Attributes:
+        model: The tree-based model (sklearn, XGBoost, LightGBM, or CatBoost)
+        feature_names: List of feature names
+        class_names: List of class names for classification
+        explainer: The underlying SHAP TreeExplainer
+        task: "classification" or "regression"
+    """
+    
+    def __init__(
+        self,
+        model,
+        feature_names: List[str],
+        class_names: Optional[List[str]] = None,
+        background_data: Optional[np.ndarray] = None,
+        task: str = "classification",
+        model_output: str = "auto",
+        feature_perturbation: str = "tree_path_dependent"
+    ):
+        """
+        Initialize the TreeSHAP explainer.
+        
+        Args:
+            model: A tree-based model or adapter containing one.
+                   Supported: RandomForest, GradientBoosting, XGBoost, 
+                   LightGBM, CatBoost, DecisionTree, ExtraTrees.
+            feature_names: List of feature names.
+            class_names: List of class names (for classification).
+            background_data: Optional background dataset for interventional
+                            feature perturbation. If None, uses tree_path_dependent.
+            task: "classification" or "regression".
+            model_output: How to transform model output. Options:
+                         - "auto": Automatically detect
+                         - "raw": Raw model output
+                         - "probability": Probability output (classification)
+                         - "log_loss": Log loss output
+            feature_perturbation: Method for handling feature perturbation:
+                                 - "tree_path_dependent": Fast, uses tree structure
+                                 - "interventional": Slower, requires background data
+        """
+        # Extract raw model if wrapped in adapter
+        raw_model = _get_raw_model(model)
+        
+        # Validate that it's a supported tree model
+        if not _is_tree_model(raw_model):
+            model_type = type(raw_model).__name__
+            raise ValueError(
+                f"TreeSHAP requires a tree-based model. Got {model_type}. "
+                f"Supported models: {', '.join(SUPPORTED_TREE_MODELS[:6])}..."
+            )
+        
+        super().__init__(model)
+        self.raw_model = raw_model
+        self.feature_names = list(feature_names)
+        self.class_names = list(class_names) if class_names else None
+        self.task = task
+        self.model_output = model_output
+        self.feature_perturbation = feature_perturbation
+        
+        # Create TreeExplainer
+        explainer_kwargs = {}
+        
+        if feature_perturbation == "interventional" and background_data is not None:
+            explainer_kwargs["data"] = background_data
+            explainer_kwargs["feature_perturbation"] = "interventional"
+        
+        if model_output != "auto":
+            explainer_kwargs["model_output"] = model_output
+        
+        self.explainer = shap.TreeExplainer(raw_model, **explainer_kwargs)
+        self.background_data = background_data
+    
+    def explain(
+        self,
+        instance: np.ndarray,
+        target_class: Optional[int] = None,
+        check_additivity: bool = False
+    ) -> Explanation:
+        """
+        Generate TreeSHAP explanation for a single instance.
+        
+        Args:
+            instance: 1D numpy array of input features.
+            target_class: For multi-class, which class to explain.
+                         If None, uses the predicted class.
+            check_additivity: Whether to verify SHAP values sum to 
+                             prediction - expected_value.
+        
+        Returns:
+            Explanation object with feature attributions.
+        """
+        instance = np.array(instance).flatten()
+        instance_2d = instance.reshape(1, -1)
+        
+        # Compute SHAP values
+        shap_values = self.explainer.shap_values(
+            instance_2d,
+            check_additivity=check_additivity
+        )
+        
+        # Handle different output formats
+        if isinstance(shap_values, list):
+            # Multi-class classification: list of arrays, one per class
+            n_classes = len(shap_values)
+            
+            if target_class is None:
+                # Use predicted class
+                if hasattr(self.raw_model, 'predict'):
+                    pred = self.raw_model.predict(instance_2d)[0]
+                    target_class = int(pred)
+                else:
+                    target_class = 0
+            
+            # Ensure target_class is valid
+            target_class = min(target_class, n_classes - 1)
+            class_shap = shap_values[target_class][0]
+            
+            # Get class name
+            if self.class_names and target_class < len(self.class_names):
+                label_name = self.class_names[target_class]
+            else:
+                label_name = f"class_{target_class}"
+            
+            # Store all class SHAP values for reference
+            all_class_shap = {
+                (self.class_names[i] if self.class_names and i < len(self.class_names) 
+                 else f"class_{i}"): shap_values[i][0].tolist()
+                for i in range(n_classes)
+            }
+        else:
+            # Binary classification or regression
+            class_shap = shap_values[0] if shap_values.ndim > 1 else shap_values.flatten()
+            label_name = self.class_names[1] if self.class_names and len(self.class_names) > 1 else "output"
+            all_class_shap = None
+        
+        # Build attributions dict
+        flat_shap = np.array(class_shap).flatten()
+        attributions = {
+            fname: float(flat_shap[i])
+            for i, fname in enumerate(self.feature_names)
+        }
+        
+        # Get expected value (base value)
+        expected_value = self.explainer.expected_value
+        if isinstance(expected_value, (list, np.ndarray)):
+            if target_class is not None and target_class < len(expected_value):
+                base_value = float(expected_value[target_class])
+            else:
+                base_value = float(expected_value[0])
+        else:
+            base_value = float(expected_value)
+        
+        explanation_data = {
+            "feature_attributions": attributions,
+            "base_value": base_value,
+            "shap_values_raw": flat_shap.tolist(),
+        }
+        
+        if all_class_shap is not None:
+            explanation_data["all_class_shap_values"] = all_class_shap
+        
+        return Explanation(
+            explainer_name="TreeSHAP",
+            target_class=label_name,
+            explanation_data=explanation_data
+        )
+    
+    def explain_batch(
+        self,
+        X: np.ndarray,
+        target_class: Optional[int] = None,
+        check_additivity: bool = False
+    ) -> List[Explanation]:
+        """
+        Generate TreeSHAP explanations for multiple instances efficiently.
+        
+        TreeSHAP can process batches more efficiently than individual calls.
+        
+        Args:
+            X: 2D numpy array of instances (n_samples, n_features).
+            target_class: For multi-class, which class to explain.
+            check_additivity: Whether to verify SHAP value additivity.
+        
+        Returns:
+            List of Explanation objects.
+        """
+        X = np.array(X)
+        if X.ndim == 1:
+            X = X.reshape(1, -1)
+        
+        # Compute SHAP values for all instances at once
+        shap_values = self.explainer.shap_values(X, check_additivity=check_additivity)
+        
+        explanations = []
+        for i in range(X.shape[0]):
+            if isinstance(shap_values, list):
+                # Multi-class
+                n_classes = len(shap_values)
+                tc = target_class if target_class is not None else 0
+                tc = min(tc, n_classes - 1)
+                class_shap = shap_values[tc][i]
+                
+                if self.class_names and tc < len(self.class_names):
+                    label_name = self.class_names[tc]
+                else:
+                    label_name = f"class_{tc}"
+            else:
+                class_shap = shap_values[i]
+                label_name = self.class_names[1] if self.class_names and len(self.class_names) > 1 else "output"
+            
+            flat_shap = np.array(class_shap).flatten()
+            attributions = {
+                fname: float(flat_shap[j])
+                for j, fname in enumerate(self.feature_names)
+            }
+            
+            expected_value = self.explainer.expected_value
+            if isinstance(expected_value, (list, np.ndarray)):
+                tc = target_class if target_class is not None else 0
+                base_value = float(expected_value[min(tc, len(expected_value) - 1)])
+            else:
+                base_value = float(expected_value)
+            
+            explanations.append(Explanation(
+                explainer_name="TreeSHAP",
+                target_class=label_name,
+                explanation_data={
+                    "feature_attributions": attributions,
+                    "base_value": base_value,
+                    "shap_values_raw": flat_shap.tolist(),
+                }
+            ))
+        
+        return explanations
+    
+    def explain_interactions(
+        self,
+        instance: np.ndarray,
+        target_class: Optional[int] = None
+    ) -> Explanation:
+        """
+        Compute SHAP interaction values for an instance.
+        
+        Interaction values show how pairs of features jointly contribute
+        to the prediction. The diagonal contains main effects.
+        
+        Args:
+            instance: 1D numpy array of input features.
+            target_class: For multi-class, which class to explain.
+        
+        Returns:
+            Explanation object with interaction matrix.
+        """
+        instance = np.array(instance).flatten()
+        instance_2d = instance.reshape(1, -1)
+        
+        # Compute interaction values
+        interaction_values = self.explainer.shap_interaction_values(instance_2d)
+        
+        # Determine target class for prediction
+        if target_class is None and hasattr(self.raw_model, 'predict'):
+            target_class = int(self.raw_model.predict(instance_2d)[0])
+        elif target_class is None:
+            target_class = 0
+        
+        # Handle different return formats from shap_interaction_values
+        if isinstance(interaction_values, list):
+            # Multi-class: list of arrays, one per class
+            n_classes = len(interaction_values)
+            tc = min(target_class, n_classes - 1)
+            interactions = np.array(interaction_values[tc][0])
+            
+            if self.class_names and tc < len(self.class_names):
+                label_name = self.class_names[tc]
+            else:
+                label_name = f"class_{tc}"
+        elif interaction_values.ndim == 4:
+            # Shape: (n_samples, n_features, n_features, n_classes)
+            n_classes = interaction_values.shape[3]
+            tc = min(target_class, n_classes - 1)
+            interactions = interaction_values[0, :, :, tc]
+            
+            if self.class_names and tc < len(self.class_names):
+                label_name = self.class_names[tc]
+            else:
+                label_name = f"class_{tc}"
+        else:
+            # Binary or regression: (n_samples, n_features, n_features)
+            interactions = interaction_values[0]
+            label_name = self.class_names[1] if self.class_names and len(self.class_names) > 1 else "output"
+        
+        # Ensure interactions is 2D (n_features x n_features)
+        interactions = np.array(interactions)
+        if interactions.ndim > 2:
+            # If still multi-dimensional, take first slice
+            interactions = interactions[:, :, 0] if interactions.ndim == 3 else interactions
+        
+        # Build interaction dict with feature name pairs
+        n_features = len(self.feature_names)
+        interaction_dict = {}
+        main_effects = {}
+        
+        for i in range(n_features):
+            fname_i = self.feature_names[i]
+            val = interactions[i, i]
+            main_effects[fname_i] = float(val) if np.isscalar(val) or val.size == 1 else float(val.flat[0])
+            
+            for j in range(i + 1, n_features):
+                fname_j = self.feature_names[j]
+                # Interaction values are symmetric, so we sum both directions
+                val_ij = interactions[i, j]
+                val_ji = interactions[j, i]
+                ij = float(val_ij) if np.isscalar(val_ij) or val_ij.size == 1 else float(val_ij.flat[0])
+                ji = float(val_ji) if np.isscalar(val_ji) or val_ji.size == 1 else float(val_ji.flat[0])
+                interaction_dict[f"{fname_i} x {fname_j}"] = ij + ji
+        
+        # Sort interactions by absolute value
+        sorted_interactions = dict(sorted(
+            interaction_dict.items(),
+            key=lambda x: abs(x[1]),
+            reverse=True
+        ))
+        
+        return Explanation(
+            explainer_name="TreeSHAP_Interactions",
+            target_class=label_name,
+            explanation_data={
+                "feature_attributions": main_effects,
+                "interactions": sorted_interactions,
+                "interaction_matrix": interactions.tolist(),
+                "feature_names": self.feature_names
+            }
+        )
+    
+    def get_expected_value(self, target_class: Optional[int] = None) -> float:
+        """
+        Get the expected (base) value of the model.
+        
+        This is the average model output over the background dataset.
+        
+        Args:
+            target_class: For multi-class, which class's expected value.
+        
+        Returns:
+            The expected value as a float.
+        """
+        expected_value = self.explainer.expected_value
+        
+        if isinstance(expected_value, (list, np.ndarray)):
+            tc = target_class if target_class is not None else 0
+            return float(expected_value[min(tc, len(expected_value) - 1)])
+        
+        return float(expected_value)