CEval 1.1.1

A framework for evaluating counterfactual explanations.

CEval — Counterfactual Explanation Evaluator

CEval is a lightweight Python package for evaluating the quality of counterfactual explanations produced by post-hoc XAI (Explainable AI) methods. It computes 14 established metrics with a single call and works with diverse architectures.

Paper: Bayrak, B., & Bach, K. (2024). Evaluation of Instance-Based Explanations: An In-Depth Analysis of Counterfactual Evaluation Metrics, Challenges, and the CEval Toolkit. IEEE Access. doi:10.1109/ACCESS.2024.3410540

---

Why CEval?

When you build or compare counterfactual explainers, you need more than one number to judge quality. CEval lets you measure all key dimensions (validity, proximity, sparsity, diversity, feasibility, and more) in a single unified framework, across different explainers and datasets.

from ceval import CEval

evaluator = CEval(samples=test_df, label="income", data=train_df, model=clf)
evaluator.add_explainer("DiCE",  dice_cfs,  "generated-cf")
evaluator.add_explainer("DICE+", dicep_cfs, "generated-cf")

print(evaluator.comparison_table)

---

Installation

pip install CEval

Requirements: Python ≥ 3.9, pandas, numpy, scikit-learn, scipy, gower, category-encoders

---

Metrics

Metric	Description	Needs model	Needs data
validity	Fraction of CFs that actually flip the classifier's prediction	✓	✓
proximity	Average feature-space distance between instance and its CF
proximity_gower	Proximity using the Gower mixed-type distance		✓
sparsity	Average fraction of features changed
count	Average number of CFs per instance
diversity	Determinant-based spread of the CF set
diversity_lcc	Diversity weighted by label-class coverage
yNN	Label consistency of the CF's k nearest neighbours	✓	✓
feasibility	Average kNN distance of CFs to the training set		✓
kNLN_dist	Distance of CF to nearest same-class neighbour		✓
relative_dist	dist(x, CF) / dist(x, NUN)		✓
redundancy	Average number of unnecessary feature changes	✓	✓
plausibility	dist(CF, NLN) / dist(NLN, NUN(NLN))	✓	✓
constraint_violation	Fraction of CFs that break user constraints

Not every metric applies to every explanation type, CEval handles this automatically and fills non-applicable cells with "-".

---

Quick Start

import pandas as pd
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split
from ceval import CEval

# 1. Prepare your data
train_df = ...   # pd.DataFrame with features + label column
test_df  = ...   # pd.DataFrame with features + label column
clf      = RandomForestClassifier().fit(train_df.drop("label", axis=1),
                                        train_df["label"])

# 2. Generate counterfactuals with your favourite explainer
#    (DiCE, PertCF, DICE, NICE, etc.)
counterfactuals = ...   # pd.DataFrame, same columns as test_df

# 3. Evaluate
evaluator = CEval(
    samples    = test_df,        # instances to explain
    label      = "label",        # target column name
    data       = train_df,       # background dataset (unlocks more metrics)
    model      = clf,            # fitted classifier (unlocks more metrics)
    k_nn       = 5,              # neighbours for kNN-based metrics
    constraints= ["age"],        # features that must not change (optional)
)

evaluator.add_explainer(
    name            = "MyExplainer",
    explanations    = counterfactuals,
    exp_type        = "generated-cf",   # "generated-cf" | "existed-cf" |
                                        # "generated-factual" | "existed-factual"
    mode            = "1to1",           # "1to1" | "1toN"
)

print(evaluator.comparison_table)

Explanation types

exp_type	When to use
"generated-cf"	Counterfactuals synthesised by an algorithm (e.g. DiCE, PertCF)
"existed-cf"	Counterfactuals retrieved from the training set
"generated-factual"	Factual explanations generated by an algorithm
"existed-factual"	Factual explanations retrieved from the training set

Explanation modes

mode	DataFrame shape	When to use
"1to1"	Same number of rows as samples	One explanation per instance
"1toN"	Any number of rows + an "instance" column	Multiple explanations per instance

---

Model Compatibility

CEval works with any classifier, not just scikit-learn.
Use the built-in wrappers from ceval.wrappers to adapt your model:

Framework	Wrapper class	Import
scikit-learn	(none needed)	pass model directly
XGBoost	XGBoostWrapper	from ceval.wrappers import XGBoostWrapper
LightGBM	LightGBMWrapper	from ceval.wrappers import LightGBMWrapper
CatBoost	CatBoostWrapper	from ceval.wrappers import CatBoostWrapper
PyTorch	TorchWrapper	from ceval.wrappers import TorchWrapper
Keras / TensorFlow	KerasWrapper	from ceval.wrappers import KerasWrapper
Anything else	GenericWrapper	from ceval.wrappers import GenericWrapper

# PyTorch
from ceval.wrappers import TorchWrapper
model = TorchWrapper(my_net, num_classes=2, device="cuda")

# XGBoost  (works with XGBClassifier and native Booster)
from ceval.wrappers import XGBoostWrapper
model = XGBoostWrapper(xgb_clf)

# Keras / TensorFlow
from ceval.wrappers import KerasWrapper
model = KerasWrapper(keras_model, num_classes=3)

# Anything else — supply two callables
from ceval.wrappers import GenericWrapper
model = GenericWrapper(
    predict_fn       = lambda X: my_model.infer(X).argmax(axis=1),
    predict_proba_fn = lambda X: my_model.infer(X),
)

# Then use as normal
evaluator = CEval(samples=test_df, label="income", data=train_df, model=model)

If you pass an incompatible model without a wrapper, CEval raises a clear TypeError that tells you exactly which wrapper to use.

---

See examples/demo_adult_income.py for a complete working demo that:

• Loads the Adult Income dataset
• Trains a Random Forest classifier
• Generates counterfactuals with DiCE
• Evaluates them in both 1-to-1 and 1-to-N mode
• Prints a full comparison table

python examples/demo_adult_income.py

Expected output:

                     DiCE (1-to-1)  DiCE (1-to-N)
validity                      0.90          0.867
proximity_gower               0.11          0.152
sparsity                      0.32          0.347
yNN                           0.68          0.713
feasibility                  48.21         183.44
redundancy                    0.80          0.733
constraint_violation          0.50          0.233
...

---

Comparing Multiple Explainers

evaluator = CEval(samples=test_df, label="label", data=train_df, model=clf)

evaluator.add_explainer("DiCE",   dice_cfs,   "generated-cf", mode="1toN")
evaluator.add_explainer("PertCF", pertcf_cfs, "generated-cf", mode="1toN")
evaluator.add_explainer("NICE",   nice_cfs,   "existed-cf",   mode="1to1")

# Side-by-side comparison
print(evaluator.comparison_table.T)

---

API Reference

CEval:

CEval(samples, label, ...)

Parameter	Type	Default	Description
samples	pd.DataFrame	required	Instances to be explained (includes label column)
label	str	required	Name of the target column
data	pd.DataFrame	None	Full background dataset; unlocks distribution-based metrics
model	sklearn estimator	None	Fitted classifier; unlocks prediction-based metrics
k_nn	int	5	Neighbours for kNN metrics
encoder	str	None	Category-encoder name for categoricals (default: OrdinalEncoder)
distance	str	None	scipy distance metric for proximity; None uses built-in mixed metric
constraints	list[str]	None	Feature names that must not change in valid CFs

evaluator.add_explainer:

evaluator.add_explainer(name, explanations, exp_type, mode="1to1")

Registers an explainer and computes all applicable metrics. Results are appended to evaluator.comparison_table.

evaluator.comparison_table:

A pd.DataFrame with one row per explainer and one column per metric. Non-applicable metrics show "-".

---

Citation

If you use CEval in your research, please cite:

@article{bayrak2024ceval,
  title   = {Evaluation of Instance-Based Explanations: An In-Depth Analysis of Counterfactual Evaluation Metrics, Challenges, and the CEval Toolkit},
  author  = {Bayrak, Bet{\"u}l and Bach, Kerstin},
  journal = {IEEE Access},
  year    = {2024},
  doi     = {10.1109/ACCESS.2024.3410540}
}

---

Related Work

This package is part of a broader research effort on counterfactual explanation methods:

• PerCE — Personalised Counterfactual Explanations (IEEE)
• PertCF — Perturbation-based Counterfactual Explainer (Paper | Code)

---

License

MIT © Betül Bayrak