Open Access

ARTICLE

Learning Scenario-Dependent Construction Strategy Selection Policies Using a Hybrid T-Spherical Fuzzy CRITIC–CoCoSo RankNet Framework

Yih-Tzoo Chen¹, Thi-Hien Dao^2,*

1 Department of Construction Engineering, College of Engineering, National Kaohsiung University of Science and Technology, Kaohsiung, Taiwan
2 Ph.D. Program in Engineering Science and Technology, College of Engineering, National Kaohsiung University of Science and Technology, Kaohsiung, Taiwan

* Corresponding Author: Thi-Hien Dao. Email:

(This article belongs to the Special Issue: Intelligent Scheduling and Optimization in Engineering and Management)

Computer Modeling in Engineering & Sciences 2026, 148(2), 27 https://doi.org/10.32604/cmes.2026.084193

Received 17 April 2026; Accepted 23 July 2026; Issue published 28 August 2026

Abstract

Construction strategy selection is a context-dependent decision problem in which multiple conflicting criteria must be considered under changing project conditions. Conventional multi-criteria decision-making (MCDM) methods generally produce rankings for predefined decision matrices but do not learn transferable preference structures across project scenarios. This study proposes a hybrid framework integrating T-spherical fuzzy sets, the Criteria Importance Through Intercriteria Correlation (CRITIC) method, the Combined Compromise Solution (CoCoSo) method, and RankNet-based pairwise learning. Fifty construction scenarios were designed using six contextual variables, and eight construction strategies were evaluated against eight criteria. Linguistic evaluations were transformed using a seven-level T-spherical fuzzy scale and the neutrality-aware score function proposed by Munir et al. Scenario-specific CRITIC–CoCoSo rankings were subsequently converted into 1400 unordered pairwise observations. To prevent information leakage and evaluate transferability, all pairs belonging to the same scenario were assigned to the same subset, with 35 scenarios used for training, seven for validation, and eight unseen scenarios for testing in each repeated split. Across ten random seeds, RankNet achieved a mean test accuracy of 0.9018 ± 0.0234, a mean Spearman rank correlation of 0.8997 ± 0.0308, and a local perturbation robustness correlation of 0.9966 ± 0.0020. Comparative analysis showed that RankNet performed comparably to logistic regression and the multilayer perceptron classifier while outperforming the examined tree-based ensemble models in ranking recovery. The ablation results further showed that interaction features improved the recovery of scenario-dependent ranking structures more clearly than direct scenario variables alone. SHAP analysis identified flexibility, risk, and their interactions with project urgency, supply instability, budget pressure, and weather uncertainty as major drivers of the learned preference policy. The proposed framework provides an interpretable and computationally efficient approach for learning construction strategy preferences under dynamic and uncertain project conditions.

---

Keywords

---