LightMIRM: Light Meta-learned Invariant Risk Minimization for Trustworthy Loan Default Prediction.

Meng Jiang,Yang Zhang,Yuan Gao,Yansong Wang,Fuli Feng,Xiangnan He

DOI: https://doi.org/10.1109/icde55515.2023.00268

2023-01-01

Abstract:Machine learning models are increasingly applied to loan default prediction to reduce the labor cost of financial institutions and the waiting time of lenders. We find that existing loan default prediction models remain lack minimax fairness, i.e., encountering significant performance drops on underrepresented subpopulations. The main cause of this trustworthy issue is pursuing Empirical Risk Minimization over the whole population, which will overlook the underrepresented subpopulations. To tackle this issue, we split the training data into subpopulations (a.k.a. environments) and conduct Invariant Risk Minimization (IRM) to learn the optimal prediction model across environments. A technical challenge is the computation cost of directly using existing IRM methods suitable for loan default prediction, such as meta-IRM, which quadratically increases as the number of environments. To reduce the complexity in training, we propose a light meta-IRM method which reduces time complexity to be linear through environment sampling and loss replaying strategies. We apply the light meta-IRM to train a representative loan default prediction model and conduct both online and offline evaluations on a large auto loan platform. Extensive experiment results validate the advantage of the proposed light meta-IRM w.r.t. the overall accuracy, minimax fairness, and training cost.