A federated learning based approach for loan defaults prediction

Abstract

The number of defaults in bank loans have recently been increasing in the past years. However, the process of sanctioning the loan has still been done manually in many of the banking organizations. Dependency on human intervention and delay in results have been the biggest obstacles in this system. While implementing machine learning models for banking applications, the security of sensitive customer banking data has always been a crucial concern and with strong legislative rules in place, sharing of data with other organizations is not possible. Along with this, the loan dataset is highly imbalanced, there are very few samples of defaults as compared to repaid loans. Hence, these problems make the default prediction system difficult to learn the patterns of defaults and thus difficult to predict them. Previous machine learning-based approaches to automate the process have been training models on the same organization's data but in today's world, classifying the loan application on the data within the organizations is no longer sufficient and a feasible solution. In this paper, we propose a federated learning-based approach for the prediction of loan applications that are less likely to be repaid which helps in resolving the above mentioned issues by sharing the weight of the model which are aggregated at the central server. The federated system is coupled with Synthetic Minority Over-sampling Technique(SMOTE) to solve the problem of imbalanced training data. Further, The federated system is coupled with a weighted aggregation based on the number of samples and performance of a worker on his dataset to further augment the performance. The improved performance by our model on publicly available real-world data further validates the same. Flexible, aggregated models can prove to be crucial in keeping out the defaulters in loan applications.