Comprehensive Root Cause Analysis of Construction Defects Using Semisupervised Graph Representation Learning

Mostofi F., Tokdemir O. B., TOĞAN V.

Journal of Construction Engineering and Management, vol.149, no.9, 2023 (SCI-Expanded) identifier

  • Publication Type: Article / Article
  • Volume: 149 Issue: 9
  • Publication Date: 2023
  • Doi Number: 10.1061/jcemd4.coeng-13435
  • Journal Name: Journal of Construction Engineering and Management
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, Aerospace Database, Applied Science & Technology Source, Business Source Elite, Business Source Premier, Communication Abstracts, Compendex, Computer & Applied Sciences, ICONDA Bibliographic, INSPEC, Metadex, Public Affairs Index, DIALNET, Civil Engineering Abstracts
  • Keywords: Association rule mining (ARM), Construction quality management (CQM), Graph representation learning (GRL), Node2vec, Root cause (RC) analysis
  • Istanbul Technical University Affiliated: Yes


Quality is a substantial pillar of construction success, as its failure poses a significant threat to the construction budget and schedule. Effective root cause (RC) analysis allows for the early identification of issues leading to quality failure and proactive defect-prevention measures. This study puts forward a flexible RC analysis method that extracts useful information from construction nonconformance reports (NCRs) to identify the future trend RCs of construction defects by employing a novel graph representation learning (GRL) approach called node2vec. Node2vec was used to connect high-cost impact RC information based on shared construction defects to determine the RCs of the construction defects. Compared with the conventional RC analysis in the literature (i.e., association rule mining), the proposed node2vec offers three advantages: (1) responsiveness to large itemset, allowing its application across multiple projects with different data collection systems. (2) It receives richer semantic information (defect-related features, RC connectivity, and different cost impacts), enabling a more comprehensive understanding of underlying defects. (3) Prediction ability of future connectivity RCs, resulting in more efficient defect-prevention actions. In contrast to unsupervised RC analysis approaches, the incorporated word2vec prediction model allows the measurement of the prediction performance of related RCs (73% accuracy and 2.31% loss), providing a noticeably more accountable RC analysis and holistic defect prevention. This in turn facilitates the integration of the proposed approach with decisions regarding quality improvement in construction projects, thereby accelerating targeted decisions and interventions within related defect-prevention policies.