Role of Shapley Additive Explanations and Resampling Algorithms for Contract Failure Prediction of Public–Private Partnership Projects

Abstract

A public–private partnership (PPP) is a common procurement model implemented worldwide as a catalyst for economic growth and improved public infrastructure. However, due to their inherent characteristics, the risk of failure in some PPP projects is high, causing heavy losses to both entities. Despite distinctive progress being made in PPP projects to reduce their failure probability, there is no proper and effective framework to predict PPP project failure in advance in either developing or in developed countries. The present study aims to develop a machine learning (ML) model to predict the failure of PPP projects to prosper in adverse conditions. This research addresses two critical issues, i.e., class imbalance and interpretability of ML models, that differentiate the current study from data-driven studies to date. First, existing studies usually focused on comparing and selecting the most adequate ML methods, but this study distinctively compared the performances of nine data resampling algorithms. Besides, in order to enhance the interpretability and visibility of the proposed model, a game theory–based feature investigation algorithm, Shapley additive explanations (SHAP), was used to identify not only the most significant features, but also the conditions of the features that cause failure or success in PPP projects. The findings illustrate that the proposed model yielded the highest prediction performance once the data set was resampled with the support vector machine-synthetic minority oversampling technique (SVM-SMOTE). SHAP analysis further shows that unsolicited proposals, domestic credit to the private sector, and project type/subtype have significant impacts on the prediction rationale. Overall, this study contributes to theory through incorporating resampling methods and SHAP algorithm into ML models as well as to practice with an advanced and reliable model to predict the status of PPP projects. The data-driven model and findings are expected to respond to current policy and industry needs by proposing a robust decision-making input for detecting risky PPP projects, allocating resources more effectively based on the most critical failure factors, and promoting the transparency of PPP project outcomes.