RealAC: A Domain-Agnostic Framework for Realistic and Actionable Counterfactual Explanations

📌 Overview

RealAC is a domain-agnostic framework for generating realistic and actionable counterfactual explanations (CFs) for machine learning models. Unlike existing methods, RealAC:

Preserves complex inter-feature dependencies without requiring explicit domain knowledge or causal graphs.

Respects user-defined feasibility constraints, enabling “frozen” features that cannot be changed.

Balances realism, actionability, and prediction validity, achieving state-of-the-art results on multiple metrics.

Counterfactual explanations describe minimal changes to input features that flip a model’s prediction. RealAC ensures these changes are plausible and implementable in practice.

🚀 Key Features

Dependency Preservation: Maintains statistical and structural integrity by minimizing divergence in mutual information between feature pairs in factual and counterfactual data.

Actionability via Binary Masking: Allows users to mark features as immutable (structural or personal preference) so they remain unchanged during optimization.

Domain-Agnostic: Works without handcrafted constraints or expert-provided causal diagrams.

Comprehensive Evaluation: Outperforms state-of-the-art and LLM-based CF generators across metrics such as:

1. Validity
2. Causal Edge Score (CES)
3. Dependency Preservation Score (DPS)
4. IM1 realism metric

📊 Data

Tested on three synthetic datasets (with sinusoidal, quadratic, exponential, and linear dependencies) and two real datasets.

1. Synthetic 1
2. Synthetic 2
3. Sangiovese
4. Diabetes
5. Adult Income

Compared against DiCE, C-CHVAE, NICE, CFNOW, CEILS, MCCE, SenseCF (LLM-based).

🔮 Key Findings:

1. RealAC consistently preserves complex non-linear dependencies better than baselines.
2. Achieves higher causal plausibility and structural integrity while maintaining competitive validity and proximity.
3. On datasets with nonlinear relationships, RealAC’s advantage is particularly strong.
4. Computational efficiency is comparable to CHVAE and faster than other methods.

📜 Citation

If you use RealAC in your research, please cite:

@inproceedings{Arefeen2025RealACAD,
  title={RealAC: A Domain-Agnostic Framework for Realistic and Actionable Counterfactual Explanations},
  author={Asiful Arefeen and Shovito Barua Soumma and Hassan Ghasemzadeh},
  year={2025}
}