Evaluation and Comparison of the Performance of Artficial Intelligence Algorithms in Predicting Construction Safety Incidents

Predicting the outcomes of safety incidents on construction projects is of a great value to various project stakeholders. Accurate estimates allow construction managers to take appropriate preventive measures based on the severity of the outcomes. Such estimates can be predicted using machine learning algorithms, although the quality of these estimates is dictated by factors including the types of algorithms employed and the dataset used to train them. Moreover, the metrics used to evaluate the algorithms can be misleading, indicating satisfactory performance when this may not be the case. In light of these considerations, this study trains a set of machine learning algorithms to predict the severity of safety incidents, highlighting the importance of confirming the credibility of performance evaluation results, and compares the performance of the algorithms. The results show that the support vector machine and k-nearest neighbors prediction models exhibit the best overall performance, with support vector machine achieving a mean absolute percentage error value of 18.78% and k-nearest neighbors an accuracy of 64.84%. On the other hand, the results also reveal that the models performed poorly in predicting some classes as a result of a high degree of imbalance identified in the dataset used for training and testing the models. The study's main contribution is to highlight the possibility of making biased performance evaluations of machine learning algorithms, depending on the performance measures used for the evaluation.