| Title |
Accident Prediction for Small-Scale Construction Sites using Administrative Data: A Cost-Sensitive DNN and Split Modeling Approach |
| Authors |
Ji-Ho Im ; Dong-Woo Lee ; Chi-Woong Moon ; In-Taek Jeon |
| DOI |
https://dx.doi.org/10.6106/KJCEM.2026.27.3.035 |
| Keywords |
Construction Accident; Risk Prediction; Split Model; Cost-Sensitive Learning |
| Abstract |
This study aims to address the data imbalance and performance degradation problems encountered in developing an AI-based construction accident risk prediction model using administrative data provided by KISCON and CSI. Existing integrated models, which were trained on the entire dataset, failed to capture the characteristics of small-scale projects, resulting in poor accident detection for sites under 500 million KRW. To overcome this limitation, this study first restored missing progress rate information by applying an S-curve regression model based on KISCON construction duration and CSI accident data, utilizing it as a key input feature. Subsequently, we proposed a separate Deep Neural Network (DNN) model specifically targeting the "under 500 million KRW" project segment. To address the severe data imbalance in small-scale projects, we implemented a cost-sensitive learning strategy by assigning higher weights to the accident class and applying decision threshold optimization. Experimental results showed that the accident detection rate (Recall) of the existing integrated model for small-scale projects was merely 1.5%. In contrast, the proposed separate model achieved a recall of 64.8%, demonstrating a significant improvement in predicting potential risks at small construction sites. These results suggest that a segregated modeling approach considering data heterogeneity by project scale is essential for building effective construction accident prevention systems. |