A Structural Topic Modeling-Based Machine Learning Approach for Pattern Extraction from Accident Data

Abstract

In occupational accident analysis, the application of machine learning is still unexplored. In addition, the presence of unstructured text data makes the analysis really difficult. Therefore, the aim of the paper is to utilize the information within accident texts using structural topic model (STM) and predict loss time injury (LTI) and non-LTI. Random forest (RF) has been used in this study for prediction as well as rule extraction. The performance of RF has also been compared with that of support vector machine (SVM), and k-nearest neighbor (KNN). The experimental results reveal that RF outperforms other classifiers in terms of accuracy. Moreover, a set of interpretable nine rules are extracted exploring the causes of the occurrence of both LTI and non-LTI.