Generational Knowledge Transfer for Model Robustness & Agility: Label Augmentation for Time-Sensitive Financial Services Applications

Abstract

Dynamic data patterns and adversarial environments pose serious challenges for machine learning models in financial service applications like financial crime and fraud detection. Adapting to continuous changes requires retraining models with the latest data patterns. This, in turn, requires balancing historical and current trends while managing training data sizes. Furthermore, the model retraining times raise problems in time-sensitive and high-volume deployment systems, in which the retraining period directly impacts the model's ability to respond to ongoing attacks in a timely manner. Achieving model agility in time-sensitive environments while staying within governance guidelines becomes a problem. In this study, we propose a temporal knowledge distillation-based label augmentation approach, which utilizes the learning from older models to rapidly boost the latest model and effectively reduces the model retraining times to achieve improved agility. Experimental results show that the proposed approach provides advantages in retraining time, while improving the model performance.