Efficient Decision Tree-Based Classification Models to Predict Safety Rating for Bridge Maintenance

Abstract

To address insufficient costs and manpower available for maintenance of aging bridges, recent research has been examining advanced maintenance technologies that can theoretically predict the condition and performance of infrastructure facilities. The current study proposes a method that is intended to predict the safety rating of bridges; among the various machine learning techniques available for this purpose, a decision tree-based classification model has been selected. Using decision tree, random forest, XGBoost (extreme gradient boosting), and LightGBM (light gradient boosting machine), 8,850 bridges on general national roads in Korea were analyzed, and the results were compared. It was possible to identify the variables that have critical impacts on the model during the model formation process. The models were analyzed through various evaluation metrics or indices such as balanced accuracy, recall, ROC (receiver operating characteristic) curve, and AUC (area under the curve). The results showed that the models using random forest, XGBoost, and LightGBM, and not those using a decision tree, exhibited excellent performance in predicting bridge safety ratings. These models achieved a recall of more than 80% for bridges with C and D ratings, which are the main targets of maintenance due to their high degree of aging. Moreover, the AUC exceeded 0.8, indicating that the prediction of bridges with ratings other than C and D was also satisfactory. These results indicate that the multiclass classification model applied in this study, with proper data sampling technique and the optimum parameters, showed improved predictive performance compared with the existing models.