In recent years, research has found that people have stable predispositions to engage in certain behavioural patterns to work safely or unsafely, which vary among individuals as a function of their personality features. In this regard, an innovative machine learning model has been recently developed to predict workers’ behavioural tendency based on personality factors. This paper presents an empirical evaluation of the model’s prediction performance (i.e. the degree to which the model can generate similar results compared to reality) to address the issue of the model’s usability before it is implemented in real situations. As virtual reality allows a good grip on fidelity resembling real-world situations, it can stimulate more natural behaviour responses from participants to increase ecological validity of experimental results. Thus, we implemented a virtual reality experimentation environment to assess workers’ safety behaviour. The model’s prediction capability was then evaluated by comparing the model prediction results and workers’ safety behaviour as assessed in virtual reality. The comparison results showed that the model predictions on two dimensions of workers’ safety behaviour (i.e. task and contextual performance) were in good agreement with the virtual reality experimental results, with Spearman correlation coefficients of 79.7% and 87.8%, respectively. The machine learning model thus proved to have good prediction capability, which allows the model to help identify vulnerable workers who are prone to undertake unsafe behaviours. The findings also suggest that virtual reality is a promising method for measuring workers’ safety behaviour as it can provide a realistic and safe environment for experimentation.

近年来，研究发现，人们有稳定的倾向来从事某些行为模式来安全或不安全地工作，这因人而异，这取决于他们的个性特征。在这方面，最近开发了一种创新的机器学习模型，以根据个性因素预测工人的行为倾向。本文提出了模型预测性能的经验评估（即模型可以生成与现实相似的结果的程度），以解决模型在实际情况中实施之前的可用性问题。由于虚拟现实可以很好地控制类似于现实世界情况的保真度，它可以激发参与者更自然的行为反应，以提高实验结果的生态有效性。因此，我们实施了一个虚拟现实实验环境来评估工人的安全行为。然后通过比较模型预测结果和虚拟现实中评估的工人安全行为来评估模型的预测能力。对比结果表明，模型对工人安全行为两个维度（即任务和情境表现）的预测与虚拟现实实验结果吻合较好，Spearman相关系数分别为79.7%和87.8%。因此，机器学习模型被证明具有良好的预测能力，这使得该模型能够帮助识别容易采取不安全行为的弱势工人。