Interpretable Machine Learning for Cost Estimation in Public Restoration Projects: An Empirical Analysis of Turkish Cultural Heritage Data

Creative Commons License

Akçakale A. H., Uğural M. N.

ADVANCES IN CIVIL ENGINEERING, cilt.2026, sa.1, ss.1-30, 2026 (SCI-Expanded, Scopus)

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 2026 Sayı: 1
  • Basım Tarihi: 2026
  • Doi Numarası: 10.1155/adce/9234704
  • Dergi Adı: ADVANCES IN CIVIL ENGINEERING
  • Derginin Tarandığı İndeksler: Applied Science & Technology Source, Middle East & Africa Database (ProQuest), Engineering Source (EBSCO), Scopus, Materials Science & Engineering Collection (ProQuest), Technology Collection (ProQuest), Science Citation Index Expanded (SCI-EXPANDED), INSPEC, Directory of Open Access Journals
  • Sayfa Sayıları: ss.1-30
  • Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu
  • İstanbul Kültür Üniversitesi Adresli: Evet

Özet

Globally, the primary barrier to adopting advanced analytics in public administration is often not a lack of data but the legal incompatibility of “black-box” predictive models with statutory requirements for transparency. This study addresses this governance challenge by developing an interpretable machine learning (ML) framework designed as an auditable solution for predicting pretender costs under strict regulatory environments, such as Turkish Public Procurement Laws No. 4734 and 5018. Utilizing a unique dataset of 578 projects from Türkiye's Directorate General of Cultural Assets and Museums, this study suggests that inherently interpretable models, specifically classification and regression trees (CART) and chi-square automatic interaction detection (CHAID), outperform traditional linear regression. Validated through 10-fold cross-validation and a temporal holdout protocol, the CART model yielded a mean absolute percentage error (MAPE) on a logarithmic scale of 8.200%, a root mean square error (RMSE) of 1.250, and a coefficient of determination (R²) of 0.663, identifying critical nonlinear risk thresholds, such as a four-fold cost escalation beyond a 100-day project duration. A key outcome was a preliminarily validated prototype decision-support tool that generated a “Decision Justification Report,” explicitly mapping project risks to auditable thresholds. This study provides a preliminarily validated, scalable framework for deploying AI in regulated public finance domains, with potential applicability to other similar regulatory contexts, helping to reconcile the tension between predictive power and institutional accountability.