SCAFFOLD-CF: Causally Constrained Counterfactual Recourse for Employee Turnover and Attrition
Counterfactual explanations identify feature changes that can flip a classifier’s prediction, but most methods treat features as independently mutable, producing recourse that conflicts with specified causal relations. This study proposes a formalized causal structure using two conditions: changes are either deliberate interventions or rule-entailed consequences, and entailed consequences are realized. Thus, SCAFFOLD-CF is proposed, which enforces both requirements as hard filters that a candidate counterfactual must respect. The main advantages of this method are the actionability and real-world alignment that a causal-aware counterfactual produces, alongside its ability to handle trade-offs with cost and manifold considerations. The method initially builds a library of candidate interventions, applies clipping, causal propagation, and snapping. It then filters candidates for predictive validity, causal admissibility, local density, and a query-scaled confidence margin before selecting diverse counterfactuals from a four-objective Pareto front. The method is evaluated against ten causally-agnostic baselines, the causally-aware ProCE, and a soft-constraint variant that penalizes causal violations as an objective, assessing cost, manifold conformity, and causal consistency. A stochastic multicriteria acceptability analysis (SMAA) aggregates three metric families and is extended with a penalty that prices abstentions. Conformity with the rule set is a property of the filter’s design. The alternatives differ: ProCE departs from the declared rules on 14.4% and 13.9% of returned counterfactuals, and the soft-constraint variant on 54.3% and 28.0%. The guarantee costs coverage, with SCAFFOLD-CF abstaining on 10.5% and 6.6% of queries. Before pricing abstentions, it is preferred under 51.5% of sampled weightings on the first dataset, 35.0 percentage points above the next method, and ranks second on attrition. Pricing abstentions reverses the attrition result, with SCAFFOLD-CF leading from a 2% penalty, and bounds the turnover result, where it leads until the penalty reaches 80% before the full-coverage causal baseline overtakes it. It is envisaged that SCAFFOLD-CF’s hard causal filter and Pareto front improve causal consistency while remaining competitive on trade-offs with other metrics, measured against an assumed rule set on employee-turnover data.
Authors
- Jia Uddin (ORCID: https://orcid.org/0000-0002-3403-4095)
- Roliana Binti Ibrahim (ORCID: https://orcid.org/0000-0001-7580-1804)
- Hasin Jawad Ali (ORCID: https://orcid.org/0009-0001-3114-8986)
- Md. Ashiful Hoque Chowdhury Tanbin
- Md. Golam Rabiul Alam
- Aryati Bakri
- Faisal Bin Al-Abid
- M. Abdullah-Al-Wadud
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-15
- DOI
- https://doi.org/10.5281/zenodo.24242054
- Primary Topic
- Explainable Artificial Intelligence (XAI)
- Type
- preprint