Deep neural networks for corporate misconduct prediction in construction industry using data from social networks

Abstract

Given the complex and continually adapting nature of misconduct, this study aims to update and interpret a misconduct prediction model using the latest techniques, i.e., deep neural networks (DNN) and social network information for Chinese construction companies. Social networks have been reported as drivers of corruption in the construction industry; however, such networks are seldom included as inputs in existing models. Based on data from 119 listed Chinese construction firms, this study applied a DNN to construct a corporate misconduct prediction model. The inputs of this model included variables related to corporate governance and financial performance, as examined in previous studies, but also social network information. This study found that the DNN demonstrated superior performance compared to benchmark models. To interpret the constructed model further and assess the role of social networks in predictions, this study also employed SHapley Additive exPlanations (SHAP), which could calculate the importance of each variable. The 20 most influential variables were identified, almost 50% of which were related to social networks. The findings indicate that the proposed model can be applied to guide shareholders and regulators in taking actions to detect and prevent corporate misconduct and aid investors in making appropriate investment decisions.