| 摘要: | 台灣工程界雖多年累積大量查核資料與標準化評分機制，但查核分數長期集
中於中高分區間，難以有效區分不同缺失型態與嚴重度等級，加上查核成績與扣點、懲罰性違約金及機關與人員考績密切相關，評分文化與制度壓力可能使分數呈現壓縮與偏態，實務上亦難以清楚清楚說明缺失嚴重程度與查核成效之量化關聯。本文所謂之查核成效，係以查核制度之評分輸出作為操作化定義，實證分析中以查核總分作為主要量化指標。
為回應此一問題，本研究以行政院公共工程委員會之公共工程標案管理系統
之資料為基礎，整併多年度、多機關及多工程類型查核事件，建構以單次查核為分析單位之資料集，並保留嚴重、中等與輕微缺失件數作為缺失嚴重度結構核心指標。方法上，結合多元線性迴歸與隨機森林等統計與機器學習工具，檢視各類缺失與查核總分之關係，並運用 K-means 分群搭配肘部法則、輪廓係數與 z 分數，自缺失結構角度辨識工程案件之典型群組與其分數分布特徵。
預期透過此一分析架構，本研究將能由制度實際運作之資料出發，釐清查核
分數在多大程度上反映缺失嚴重程度，揭示制度設計盲點與改善方向，並協助主管機關將既有查核資料轉化為可支援缺失嚴重度分群與資源配置之決策基礎，同時提供承攬與監造單位進行自我診斷與品質精進之量化依據。;Taiwan’s construction sector has accumulated extensive inspection data and standardized scoring mechanisms, yet scores remain highly concentrated in the mid to high range. As inspection results are tied to demerit points, punitive liquidated damages, and performance evaluations, institutional and cultural pressures may compress and skew score distributions, making it unclear to what extent inspection scores truly reflect defect patterns and severity. In this study, “audit effectiveness” is operationalized as the scoring output of the public construction inspection system. Accordingly, the overall inspection score is adopted as the primary quantitative indicator in the empirical analyses.
To address this issue, this study uses data from the Public Construction Bidding and Management System of the Public Construction Commission, constructing a dataset at the level of individual inspection events and retaining counts of severe, moderate, and minor defects as core indicators of defect severity. Multiple linear regression and random forest models are applied to examine the relationships between these defect categories and the normalized overall inspection score. K-means clustering, combined with the elbow method, silhouette coefficient, and z-score analysis, is then used to identify typical defect-structure clusters and their associated score patterns. The proposed framework aims to clarify how well inspection scores reflect defect structures, reveal blind spots in the current scoring design, and provide a basis for more informed defect severity, resource allocation, and data-driven quality improvement for both authorities and practitioners. |
