An improved light gradient boosting machine algorithm based on swarm algorithms for predicting loan default of peer-to-peer lending

Abstract

Internet finance and big data technology are booming in the world. The launch of peer to peer (P2P) lending platforms is a sign and a great opportunity for entrepreneurs to easily increase their capital injection. However, this great opportunity has a high risk of impacting the sustainability and security development of the platform. One way to minimize loan risk is to predict the possibility of loan default. Hence, this study aims to find the best predictive model for predicting loan default of P2P Lending Club dataset. An improved light gradient boosting machine (LightGBM) via features selection by using swarm algorithms i.e. Ant colony optimization (ACO) and bee colony optimization (BCO) to the prediction analysis process. The best feature selection process is selected 6 out of 18 features. The synthetic minority oversampling technique (SMOTE) method is also provided to solve the unbalance class problem in the dataset, then a series of operations such as data cleaning and dimension reduction are performed. The experimental results prove that the LightGBM algorithm has been successfully improved. This success is shown by the prediction accuracy of LightGBM+ACO is 95.64%, LighGBM+BCO is 94.70% and LightGBM is 94.38%. This success also demonstrates outstanding performance in predicting loan default and strong generalizations.