| Description: |
I. BACKGROUND
Counterfactual explanations ask what would need to change for a particular instance to obtain a different outcome. Standard methods often search for nearby feature values that reverse the prediction of a machine-learning classifier, such as DiCE studies
proximity, diversity, and feasibility of alternative examples [1]. However, changing a feature in a dataset is not necessarily an intervention that could be performed in the real world. Structural causal models (SCMs) provide a way to represent the
consequences of interventions and distinguish a predicted change from a causally meaningful one [2]. Reliability remains challenging when the underlying SCM is learned rather than known. Observational data may be compatible
with several causal mechanisms, particularly with few observations or rare outcomes. Prior work has demonstrated the difficulty of guaranteeing recourse under imperfect causal knowledge [3]. Recent research also identifies several forms of counterfactual
robustness, including sensitivity to model changes and input perturbations [4], [5]. These findings motivate a focused assessment of how uncertainty in an assumed causal model affects the credibility of individual counterfactual explanations.
II. PROBLEM STATEMENT
Counterfactual explanations are frequently judged by whether they change a model’s prediction and require only small input modifications. These properties do not establish that the proposed changes are feasible, causally valid, or stable across plausible
causal models. With limited observations and imbalanced outcomes, learned mechanisms may be especially uncertain. The research problem is therefore to determine when a counterfactual explanation can be trusted, and how limitations of the data
and causal assumptions should qualify its interpretation.
III. AIM
To assess and characterize the reliability of counterfactual explanations generated from learned causal models under data scarcity, class imbalance, and causal-model uncertainty.
IV. OBJECTIVES
1) Examine the effects of sample size and minority-class representation on counterfactual quality.
2) Assess the sensitivity of counterfactual explanations to plausible changes in learned causal mechanisms.
3) Identify measurable conditions for distinguishing robust explanations from unreliable or unsupported ones.
V. RESEARCH QUESTIONS
RQ1: How do data scarcity and class imbalance affect counterfactual reliability?
This question investigates how limited observations and minority-class representation influence the validity, feasibility, and
stability of counterfactual explanations.
RQ2: How does causal-model uncertainty affect counterfactual explanations?
This question examines how variations in learned causal mechanisms influence counterfactual outcomes and the consistency of
recommended interventions.
RQ3: When can counterfactual explanations be considered reliable?
This question aims to identify measurable criteria for determining whether counterfactual explanations are robust, feasible, and
sufficiently supported by the available causal evidence.
VI. EXPECTED OUTCOMES
• Understanding the impact of data limitations: An empirical assessment of how sample size and class imbalance affect
counterfactual validity, feasibility, and stability, particularly for minority-class instances.
• Characterization of causal-model uncertainty: A quantitative analysis of how uncertainty in learned causal models
influences counterfactual outcomes and recommended interventions, identifying explanations that remain consistent across
plausible causal models.
• Criteria for counterfactual reliability: A set of measurable criteria for assessing the reliability of counterfactual
explanations and identifying cases where explanations are unstable, infeasible, or insufficiently supported by the available
evidence.
REFERENCES
[1] R. K. Mothilal, A. Sharma, and C. Tan, “Explaining machine learning classifiers through diverse counterfactual explanations,” in Proceedings of the 2020
Conference on Fairness, Accountability, and Transparency. ACM, 2020, pp. 607–617.
[2] A.-H. Karimi, B. Schölkopf, and I. Valera, “Algorithmic recourse: from counterfactual explanations to interventions,” in Proceedings of the 2021 ACM
Conference on Fairness, Accountability, and Transparency. ACM, 2021, pp. 353–362.
[3] A.-H. Karimi, J. von Kügelgen, B. Schölkopf, and I. Valera, “Algorithmic recourse under imperfect causal knowledge: A probabilistic approach,” in
Advances in Neural Information Processing Systems, vol. 33, 2020.
[4] J. Jiang, F. Leofante, A. Rago, and F. Toni, “Robust counterfactual explanations in machine learning: A survey,” in Proceedings of the Thirty-Third
International Joint Conference on Artificial Intelligence, IJCAI-24, K. Larson, Ed. International Joint Conferences on Artificial Intelligence Organization,
8 2024, pp. 8086–8094, survey Track. [Online]. Available: https://doi.org/10.24963/ijcai.2024/894
[5] L. Marzari, F. Leofante, F. Cicalese, and A. Farinelli, “Probabilistically robust counterfactual explanations under model changes,” Artificial Intelligence, p.
104459, 2025
|
