Performance and Key Variable Analysis of Construction Accident Prediction Model Incorporating Periodic Features

초록

The Construction Safety Management Integrated Information system (CSI), operated by the Korea Authority of Land & Infrastructure Safety (KALIS) since 2019, provides a comprehensive database of construction accident records to support data-driven safety management. This resource has enabled the development of prediction models aimed at preventing accidents and enhancing decision-making on construction sites. While these models have shown promising performance, opportunities remain to improve their accuracy by considering the periodic nature of the CSI, which is updated regularly with new accident cases. This study investigates how the duration of training data affects the performance of construction accident prediction models when working with periodically updated datasets. To explore this, multiple training datasets were constructed using different durations of data using a total of 29,151 accident records collected between 2019 and 2024. All models were developed using the eXtreme Gradient Boosting (XGBoost) algorithm, a widely adopted method in construction safety prediction. Model performance was evaluated using the F1-score to ensure fair and consistent comparison. Furthermore, SHapley Additive exPlanations (SHAP) analysis was employed to identify key predictive features and examine how their importance varied with different training periods. The findings revealed that training models on the full historical dataset does not always lead to the most accurate predictive performance. In many cases, models trained on selectively chosen periods outperformed those using the entire dataset. The primary contribution of this study lies in proposing a systematic framework for optimizing model performance by aligning training data with periodic characteristics without enlarging the data size. © 2025, Korea Institute of Construction Engineering and Management. All rights reserved.

키워드