Public construction quality in Taiwan is commonly assessed through committee-based inspections, yet the resulting scores are often subjective and heavily concentrated within narrow grading ranges. To address this limitation, this study proposes a data-driven framework that integrates Principal Component Analysis (PCA) with a Multilayer Neural Network (MNN) to reconstruct objective and discriminative ranges of project quality scores. Using 962 inspection records from the Public Construction Intelligence Cloud (PCIC), PCA is first applied to reduce 499 defect items into 13 representative serious defects, mitigating multicollinearity and retaining the most informative quality indicators. These defects, together with the project contract amount and construction progress, are then used as inputs to an optimized MNN classifier. A systematic hyperparameter search and stratified 10-fold cross-validation are employed to ensure robust model generalization. Based on the learned relationships, new grading thresholds are derived: A+ (86–100), A (83–86), A– (80–83), and B+ (<80). The proposed PCA–MNN framework achieves an overall accuracy of 95% and significantly alleviates the extreme class imbalance observed in the original scoring scheme. Results demonstrate that the reconstructed ranges provide a more balanced, interpretable, and objective representation of project quality, enabling fairer multi-class evaluation and supporting more reliable decision-making in public construction quality management.
Ching-Lung Fan (Tue,) studied this question.