CounterEval: Towards Unifying Evaluation of Counterfactual Explanations

Overview

CounterEval offers a human-evaluated dataset of 30 counterfactual scenarios, engineered to serve as a benchmark for evaluating a wide range of counterfactual explanation generation frameworks. Each scenario has been strategically designed to exhibit varying levels across multiple explanatory quality metrics, such as Feasibility, Consistency, Trust, Completeness, Fairness, and Complexity, while maintaining high Understandability and using Overall Satisfaction as a general measure of quality. The scenarios are based on well-known datasets like the Adult dataset and the Pima Indians Diabetes dataset, which include a mix of categorical and continuous data. Additional fabricated scenarios were created to explore extreme values in certain metrics. Each scenario has been evaluated by 206 individuals recruited through the Prolific platform, across all metrics, enabling detailed analysis of how individual explanatory metrics influence the overall satisfaction score.

Evaluation Metrics

Below is the detailed table of metrics with their definitions and scales:

Metric	Definition	Scale
Overall Satisfaction	This scenario effectively explains how to reach a different outcome.	from 1 (very unsatisfied) to 6 (very satisfied)
Feasibility	The actions suggested by the explanation are practical, realistic to implement, and actionable.	from 1 (very infeasible) to 6 (very easy to do)
Consistency	All parts of the explanation are logically coherent and do not contradict each other.	from 1 (very inconsistent) to 6 (very consistent)
Completeness	The explanation is sufficient in explaining the outcome.	from 1 (very incomplete) to 6 (very complete)
Trust	I believe that the suggested changes would bring about the desired outcome.	from 1 (not at all) to 6 (very much)
Understandability	I feel like I understood the phrasing of the explanation well.	from 1 (incomprehensible) to 6 (very understandable)
Fairness	The explanation is unbiased towards different user groups and does not operate on sensitive features.	from 1 (very biased) to 6 (completely fair)
Complexity	The explanation has an appropriate level of detail and complexity - not too simple, yet not overly complex.	from -2 (too simple) to 0 (ideal complexity) to 2 (too complex)

Evaluators

The evaluators were recruited through the Prolific platform, with an average survey completion time of 42 minutes. The selection was based on the criteria of being at least 18 years old, fluent in English. Here’s an overview of the evaluator demographics and their distribution:

Demographic	Description	Number	Percentage (%)
Age Group
	18-24 years old	64	31.07
	25-34 years old	95	46.12
	35-44 years old	25	12.14
	45-54 years old	15	7.28
	55-64 years old	7	3.40
Nationality
	South African	32	15.53
	Polish	30	14.56
	British	27	13.11
	Mexican	20	9.71
	Portuguese	19	9.22
	Chilean	12	5.83
	Italian	8	3.88
	Greek	7	3.40
	Hungarian	7	3.40
	USA	6	2.91
	Other	38	18.45
Education Level
	Primary school or lower	1	0.49
	High school	69	33.50
	Bachelor's degree or equivalent	93	45.15
	Master's degree or equivalent	39	18.93
	Doctoral degree or equivalent	4	1.94
English Proficiency
	Native speaker / Fully proficient	166	80.58
	Moderately proficient	40	19.42
Experience in Machine Learning
	No experience	100	48.54
	Some experience	89	43.20
	I am studying in a related field	12	5.83
	I work in the field / Extensive experience	5	2.43
Experience with Causality Frameworks
	No	145	70.39
	I am familiar with the general concept	57	27.67
	I have previous experience with them	4	1.94
Medical Background
	No	181	87.86
	I am currently studying medicine or a related field	10	4.85
	I work with medical data or in a field related to medicine	8	3.88
	I have a degree in medicine or a related field	4	1.94
	I work in the field of medicine	3	1.46

Exclusion criteria

The responses from participants were excluded based on the following quality checks:

• Attention Check: Participants were required to write specific text in a field, differing from the usual question format, to ensure attentiveness.
• Response Time: Participants who completed the survey significantly faster than the expected minimum time were flagged.
• Average Understandability Time: Participants with consistently low understandability ratings were flagged.
• Indicator questions answerd: Specific questions with known extreme values for certain criteria were included to verify if participants' answers aligned with the expected evaluation. For example, marking infeasible scenarios (e.g., "change place of birth") as feasible.
• Uniform Response Patterns: Participants providing uniform or highly unvaried ratings across all metrics.

Respondents failing in 3 aforementioned criteria were removed. IDs of excluded participants: 12, 27, 28, 44, 75, 83, 84, 88, 97, 201.

File Descriptions

• cleaned_cf_dataset.csv: This file contains the cleaned version of evaluator responses after applying exclusion criteria. Each row corresponds to a Scenario-Metric pair (30 scenarios × 8 metrics = 240 rows). Each column corresponds to a participant's ranking of the given Scenario-Metric pair.
• full_cf_dataset.csv: This file contains all evaluator responses before applying exclusion criteria.
• survey_question.csv: This file provides a list of all unique counterfactual scenarios used in the study, without duplicates.
• participant_background.csv: This file provides background information on the participants who evaluated the counterfactual scenarios. It includes demographic details such as age, nationality, education level, and expertise in related fields.

Acknowledgments

This research was supported by the European Union’s Horizon 2020 Research and Innovation Programme under Grant Agreement No. 952060 (Trust AI), and by the Estonian Centre of Excellence in Artificial Intelligence (EXAI), funded by the Estonian Ministry of Education and Research grant TK213, Estonian Research Council Grants PRG1604, PSG728 and EKTB55 project.

Citation

Please cite our work using the following format if you use this dataset:

@misc{domnich2024unifyingevaluationcounterfactualexplanations,
      title={Towards Unifying Evaluation of Counterfactual Explanations: Leveraging Large Language Models for Human-Centric Assessments},
      author={Marharyta Domnich and Julius Valja and Rasmus Moorits Veski and Giacomo Magnifico and Kadi Tulver and Eduard Barbu and Raul Vicente},
      year={2024},
      eprint={2410.21131},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2410.21131},
}