import pandas as pd
from rocelib.evaluations.robustness_evaluations.Evaluator import Evaluator
[docs]
class ValidityEvaluator(Evaluator):
"""
An Evaluator class which evaluates the proportion of recourses which are valid
...
Attributes / Properties
-------
task: Task
Stores the Task for which we are evaluating the validity of CEs
-------
Methods
-------
evaluate() -> int:
Returns the proportion of CEs which are valid
-------
"""
[docs]
def checkValidity(self, instance, valid_val):
"""
Checks if a given CE is valid
@param instance: pd.DataFrame / pd.Series / torch.Tensor, the CE to check validity of
@param valid_val: int, the target column value which denotes a valid CE
@return:
"""
# Convert to DataFrame if needed
if not isinstance(instance, pd.DataFrame):
instance = pd.DataFrame(instance).T
# Drop categorical or non-numeric columns before prediction
instance = instance.drop(columns=["predicted", "loss", "target"], errors="ignore")
# Align instance columns to match model expected features
expected_features = self.task.dataset.X.columns
instance = instance[expected_features] # Select only expected features
# Convert to float32 to avoid numpy.object_ errors
instance = instance.astype("float32")
# Ensure instance is not empty
if instance.empty:
return False
# Predict and check validity
return self.task.model.predict_single(instance) == valid_val
[docs]
def evaluate(self, recourse_method, valid_val=1, column_name="target", **kwargs):
"""
Evaluates the proportion of CEs are valid
@param recourses: pd.DataFrame, set of CEs which we want to evaluate
@param valid_val: int, target column value which denotes a valid instance
@param column_name: str, name of target column
@param kwargs: other arguments
@return: int, proportion of CEs which are valid
"""
recourses = self.task._CEs[recourse_method][0]
valid = 0
cnt = 0
# Remove redundant columns
instances = recourses.drop(columns=[column_name, "loss"], errors='ignore')
for _, instance in instances.iterrows():
# Increment validity counter if CE is valid
if instance is not None and not instance.empty and self.checkValidity(instance, valid_val):
valid += 1
# Increment total counter
cnt += 1
return valid / cnt