Skip to content

Каузальный анализ на основе вмешательств

Ссылка на руководство пользователя

Source code in applybn/explainable/causal_analysis/intervention_causal_effect.py 
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
class InterventionCausalExplainer:
    def __init__(self, n_estimators=10):
        """Initialize the ModelInterpreter.

        Attributes:
            n_estimators: Number of estimators for Data-IQ.
        """
        logger.info(
            "Initializing InterventionCausalExplainer with %d estimators", n_estimators
        )
        self.n_estimators = n_estimators
        self.clf = None
        self.dataiq_train = None
        self.dataiq_test = None
        self.confidence_train = None
        self.confidence_test = None
        self.aleatoric_uncertainty_train = None
        self.aleatoric_uncertainty_test = None
        self.feature_effects = None
        self.confidence_test_before_intervention = None
        self.aleatoric_uncertainty_test_before_intervention = None

    def train_model(self, model: BaseEstimator | ClassifierMixin, X, y):
        """Train the model on the training data.

        Args:
            model: The model to train
            X: Training data
            y: Training labels
        """
        logger.info("Training the model with %d samples", X.shape[0])
        self.clf = model
        self.clf.fit(X, y)
        logger.info("Model training complete.")

    def _compute_confidence_uncertainty(self, X, y, suffix: str):
        """Helper to compute confidence and uncertainty for train/test data using Data-IQ."""
        data_type = "training" if suffix == "train" else "test"
        logger.info(
            f"Computing confidence and uncertainty on {data_type} data using Data-IQ."
        )

        dataiq = DataIQSKLearn(X=X, y=y)
        dataiq.on_epoch_end(clf=self.clf, iteration=self.n_estimators)

        setattr(self, f"dataiq_{suffix}", dataiq)
        setattr(self, f"confidence_{suffix}", dataiq.confidence)
        setattr(self, f"aleatoric_uncertainty_{suffix}", dataiq.aleatoric)

    def compute_confidence_uncertainty_train(self, X, y):
        """Compute model confidence and aleatoric uncertainty on training data using Data-IQ."""
        self._compute_confidence_uncertainty(X, y, "train")

    def compute_confidence_uncertainty_test(self, X, y):
        """Compute model confidence and aleatoric uncertainty on test data using Data-IQ."""
        self._compute_confidence_uncertainty(X, y, "test")

    def estimate_feature_impact(self, X, random_state=42):
        """Estimate the causal effect of each feature on the model's confidence using training data."""
        logger.info(
            "Estimating feature impact using causal inference on training data."
        )
        self.feature_effects = {}
        for feature in track(X.columns, description="Estimating feature impacts"):
            logger.debug(f"Estimating effect of feature '{feature}'.")
            treatment = X[feature].values
            outcome = self.confidence_train
            covariates = X.drop(columns=[feature])

            est = CausalForestDML(
                model_y=RandomForestRegressor(),
                model_t=RandomForestRegressor(),
                discrete_treatment=False,
                random_state=random_state,
            )
            est.fit(Y=outcome, T=treatment, X=covariates)
            te = est.const_marginal_effect(covariates).mean()
            self.feature_effects[feature] = te

        # Convert to Series and sort
        self.feature_effects = (
            pd.Series(self.feature_effects).abs().sort_values(ascending=False)
        )
        logger.info("Feature effects estimated.")

    def plot_aleatoric_uncertainty(self, before_intervention: bool = True):
        """Plot aleatoric uncertainty for test data before and after intervention."""
        if before_intervention:
            plt.figure(figsize=(10, 5))
            plt.hist(
                self.aleatoric_uncertainty_test_before_intervention,
                bins=30,
                alpha=0.5,
                label="Uncertainty Before Intervention",
            )
            plt.hist(
                self.aleatoric_uncertainty_test,
                bins=30,
                alpha=0.5,
                color="red",
                label="Uncertainty After Intervention",
            )
            plt.title("Test Data: Aleatoric Uncertainty Before and After Intervention")
            plt.xlabel("Aleatoric Uncertainty")
            plt.ylabel("Frequency")
            plt.legend()
            plt.show()

    def plot_top_feature_effects(self, top_n: int = 10):
        """Plot a bin plot of the top N most impactful features with their causal effects.

        Args:
            top_n: Number of top features to plot.
        """
        top_features = self.feature_effects.head(top_n)
        plt.figure(figsize=(10, 8))
        top_features.plot(kind="bar", color="skyblue")
        plt.title(f"Top {top_n} Most Impactful Features by Causal Effect")
        plt.xlabel("Features")
        plt.ylabel("Causal Effect")
        plt.xticks(rotation=45, ha="right")
        plt.tight_layout()
        plt.show()

    def perform_intervention(self, X_test, y_test):
        """Perform an intervention on the top 5 most impactful features in the test data and observe changes."""
        if self.feature_effects is None:
            raise ValueError("Feature effects have not been estimated yet.")

        top_features = self.feature_effects.head(5).index.tolist()
        logger.info(f"Top {len(top_features)} most impactful features: {top_features}")

        # Compute confidence on test data before intervention
        self.compute_confidence_uncertainty_test(X=X_test, y=y_test)
        self.confidence_test_before_intervention = self.confidence_test.copy()
        self.aleatoric_uncertainty_test_before_intervention = (
            self.aleatoric_uncertainty_test.copy()
        )

        original_feature_values_test = X_test[top_features].copy()

        for feature in track(top_features, description="Performing interventions"):
            plt.figure(figsize=(10, 5))
            plt.hist(
                original_feature_values_test[feature],
                bins=30,
                alpha=0.5,
                label="Before Intervention",
            )

            logger.debug(f"Performing intervention on '{feature}' in test data.")
            min_val = original_feature_values_test[feature].min()
            max_val = original_feature_values_test[feature].max()
            np.random.seed(42)
            new_values = np.random.uniform(
                low=min_val, high=max_val, size=X_test.shape[0]
            )
            X_test[feature] = new_values

            plt.hist(
                X_test[feature],
                bins=30,
                alpha=0.5,
                color="orange",
                label="After Intervention",
            )
            plt.title(
                f"Test Data: Distribution of '{feature}' Before and After Intervention"
            )
            plt.xlabel(feature)
            plt.ylabel("Frequency")
            plt.legend()
            plt.show()

        self.compute_confidence_uncertainty_test(X=X_test, y=y_test)

        plt.figure(figsize=(10, 5))
        plt.hist(
            self.confidence_test_before_intervention,
            bins=30,
            alpha=0.5,
            label="Confidence Before Intervention",
        )
        plt.hist(
            self.confidence_test,
            bins=30,
            alpha=0.5,
            color="green",
            label="Confidence After Intervention",
        )
        plt.title(
            f"Test Data: Model Confidence Before and After Intervention on {len(top_features)} features"
        )
        plt.xlabel("Confidence")
        plt.ylabel("Frequency")
        plt.legend()
        plt.show()

        self.plot_aleatoric_uncertainty()

        logger.info(
            "Intervention complete. Observed changes in model confidence on test data."
        )

    def interpret(
        self,
        model,
        X_train: pd.DataFrame,
        y_train: pd.Series,
        X_test: pd.DataFrame,
        y_test: pd.Series,
    ):
        """Run the full interpretation process."""
        self.train_model(model=model, X=X_train, y=y_train)
        self.compute_confidence_uncertainty_train(X=X_train, y=y_train)
        self.estimate_feature_impact(X=X_train)
        self.plot_top_feature_effects()
        self.perform_intervention(X_test=X_test, y_test=y_test)

__init__(n_estimators=10)

Initialize the ModelInterpreter.

Attributes:

Name Type Description
n_estimators

Number of estimators for Data-IQ.
Source code in applybn/explainable/causal_analysis/intervention_causal_effect.py 
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
def __init__(self, n_estimators=10):
    """Initialize the ModelInterpreter.

    Attributes:
        n_estimators: Number of estimators for Data-IQ.
    """
    logger.info(
        "Initializing InterventionCausalExplainer with %d estimators", n_estimators
    )
    self.n_estimators = n_estimators
    self.clf = None
    self.dataiq_train = None
    self.dataiq_test = None
    self.confidence_train = None
    self.confidence_test = None
    self.aleatoric_uncertainty_train = None
    self.aleatoric_uncertainty_test = None
    self.feature_effects = None
    self.confidence_test_before_intervention = None
    self.aleatoric_uncertainty_test_before_intervention = None

compute_confidence_uncertainty_test(X, y)

Compute model confidence and aleatoric uncertainty on test data using Data-IQ.

Source code in applybn/explainable/causal_analysis/intervention_causal_effect.py 
71
72
73
def compute_confidence_uncertainty_test(self, X, y):
    """Compute model confidence and aleatoric uncertainty on test data using Data-IQ."""
    self._compute_confidence_uncertainty(X, y, "test")

compute_confidence_uncertainty_train(X, y)

Compute model confidence and aleatoric uncertainty on training data using Data-IQ.

Source code in applybn/explainable/causal_analysis/intervention_causal_effect.py 
67
68
69
def compute_confidence_uncertainty_train(self, X, y):
    """Compute model confidence and aleatoric uncertainty on training data using Data-IQ."""
    self._compute_confidence_uncertainty(X, y, "train")

estimate_feature_impact(X, random_state=42)

Estimate the causal effect of each feature on the model's confidence using training data.

Source code in applybn/explainable/causal_analysis/intervention_causal_effect.py 
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
def estimate_feature_impact(self, X, random_state=42):
    """Estimate the causal effect of each feature on the model's confidence using training data."""
    logger.info(
        "Estimating feature impact using causal inference on training data."
    )
    self.feature_effects = {}
    for feature in track(X.columns, description="Estimating feature impacts"):
        logger.debug(f"Estimating effect of feature '{feature}'.")
        treatment = X[feature].values
        outcome = self.confidence_train
        covariates = X.drop(columns=[feature])

        est = CausalForestDML(
            model_y=RandomForestRegressor(),
            model_t=RandomForestRegressor(),
            discrete_treatment=False,
            random_state=random_state,
        )
        est.fit(Y=outcome, T=treatment, X=covariates)
        te = est.const_marginal_effect(covariates).mean()
        self.feature_effects[feature] = te

    # Convert to Series and sort
    self.feature_effects = (
        pd.Series(self.feature_effects).abs().sort_values(ascending=False)
    )
    logger.info("Feature effects estimated.")

interpret(model, X_train, y_train, X_test, y_test)

Run the full interpretation process.

Source code in applybn/explainable/causal_analysis/intervention_causal_effect.py 
222
223
224
225
226
227
228
229
230
231
232
233
234
235
def interpret(
    self,
    model,
    X_train: pd.DataFrame,
    y_train: pd.Series,
    X_test: pd.DataFrame,
    y_test: pd.Series,
):
    """Run the full interpretation process."""
    self.train_model(model=model, X=X_train, y=y_train)
    self.compute_confidence_uncertainty_train(X=X_train, y=y_train)
    self.estimate_feature_impact(X=X_train)
    self.plot_top_feature_effects()
    self.perform_intervention(X_test=X_test, y_test=y_test)

perform_intervention(X_test, y_test)

Perform an intervention on the top 5 most impactful features in the test data and observe changes.

Source code in applybn/explainable/causal_analysis/intervention_causal_effect.py 
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
def perform_intervention(self, X_test, y_test):
    """Perform an intervention on the top 5 most impactful features in the test data and observe changes."""
    if self.feature_effects is None:
        raise ValueError("Feature effects have not been estimated yet.")

    top_features = self.feature_effects.head(5).index.tolist()
    logger.info(f"Top {len(top_features)} most impactful features: {top_features}")

    # Compute confidence on test data before intervention
    self.compute_confidence_uncertainty_test(X=X_test, y=y_test)
    self.confidence_test_before_intervention = self.confidence_test.copy()
    self.aleatoric_uncertainty_test_before_intervention = (
        self.aleatoric_uncertainty_test.copy()
    )

    original_feature_values_test = X_test[top_features].copy()

    for feature in track(top_features, description="Performing interventions"):
        plt.figure(figsize=(10, 5))
        plt.hist(
            original_feature_values_test[feature],
            bins=30,
            alpha=0.5,
            label="Before Intervention",
        )

        logger.debug(f"Performing intervention on '{feature}' in test data.")
        min_val = original_feature_values_test[feature].min()
        max_val = original_feature_values_test[feature].max()
        np.random.seed(42)
        new_values = np.random.uniform(
            low=min_val, high=max_val, size=X_test.shape[0]
        )
        X_test[feature] = new_values

        plt.hist(
            X_test[feature],
            bins=30,
            alpha=0.5,
            color="orange",
            label="After Intervention",
        )
        plt.title(
            f"Test Data: Distribution of '{feature}' Before and After Intervention"
        )
        plt.xlabel(feature)
        plt.ylabel("Frequency")
        plt.legend()
        plt.show()

    self.compute_confidence_uncertainty_test(X=X_test, y=y_test)

    plt.figure(figsize=(10, 5))
    plt.hist(
        self.confidence_test_before_intervention,
        bins=30,
        alpha=0.5,
        label="Confidence Before Intervention",
    )
    plt.hist(
        self.confidence_test,
        bins=30,
        alpha=0.5,
        color="green",
        label="Confidence After Intervention",
    )
    plt.title(
        f"Test Data: Model Confidence Before and After Intervention on {len(top_features)} features"
    )
    plt.xlabel("Confidence")
    plt.ylabel("Frequency")
    plt.legend()
    plt.show()

    self.plot_aleatoric_uncertainty()

    logger.info(
        "Intervention complete. Observed changes in model confidence on test data."
    )

plot_aleatoric_uncertainty(before_intervention=True)

Plot aleatoric uncertainty for test data before and after intervention.

Source code in applybn/explainable/causal_analysis/intervention_causal_effect.py 
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
def plot_aleatoric_uncertainty(self, before_intervention: bool = True):
    """Plot aleatoric uncertainty for test data before and after intervention."""
    if before_intervention:
        plt.figure(figsize=(10, 5))
        plt.hist(
            self.aleatoric_uncertainty_test_before_intervention,
            bins=30,
            alpha=0.5,
            label="Uncertainty Before Intervention",
        )
        plt.hist(
            self.aleatoric_uncertainty_test,
            bins=30,
            alpha=0.5,
            color="red",
            label="Uncertainty After Intervention",
        )
        plt.title("Test Data: Aleatoric Uncertainty Before and After Intervention")
        plt.xlabel("Aleatoric Uncertainty")
        plt.ylabel("Frequency")
        plt.legend()
        plt.show()

plot_top_feature_effects(top_n=10)

Plot a bin plot of the top N most impactful features with their causal effects.

Parameters:

Name Type Description Default
top_n int

Number of top features to plot.

 10
Source code in applybn/explainable/causal_analysis/intervention_causal_effect.py 
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
def plot_top_feature_effects(self, top_n: int = 10):
    """Plot a bin plot of the top N most impactful features with their causal effects.

    Args:
        top_n: Number of top features to plot.
    """
    top_features = self.feature_effects.head(top_n)
    plt.figure(figsize=(10, 8))
    top_features.plot(kind="bar", color="skyblue")
    plt.title(f"Top {top_n} Most Impactful Features by Causal Effect")
    plt.xlabel("Features")
    plt.ylabel("Causal Effect")
    plt.xticks(rotation=45, ha="right")
    plt.tight_layout()
    plt.show()

train_model(model, X, y)

Train the model on the training data.

Parameters:

Name Type Description Default
model BaseEstimator | ClassifierMixin

The model to train

 required
X

Training data

 required
y

Training labels

 required
Source code in applybn/explainable/causal_analysis/intervention_causal_effect.py 
40
41
42
43
44
45
46
47
48
49
50
51
def train_model(self, model: BaseEstimator | ClassifierMixin, X, y):
    """Train the model on the training data.

    Args:
        model: The model to train
        X: Training data
        y: Training labels
    """
    logger.info("Training the model with %d samples", X.shape[0])
    self.clf = model
    self.clf.fit(X, y)
    logger.info("Model training complete.")

Пример

from applybn.explainable.causal_analysis import InterventionCausalExplainer
from sklearn.ensemble import RandomForestClassifier

interpreter = InterventionCausalExplainer()
interpreter.interpret(RandomForestClassifier(), X_train, y_train, X_test, y_test)