Improving Drift Detection by Monitoring Shapley Loss Values

Bastien Zimmermann, Matthieu Boussard

Abstract: Along the deployment of Machine Learning models rises an inherent need for monitoring, where model performances should be tracked as well as potential drifts. In a live environment, with evolving data, the risk is for the model to become ill-adapted for the given situation. The failure to detect drift while leading to a performance deterioration could also cause side effects due to model over-trust. Informing the user of any anomaly upon detection is the key to enabling any action. We propose Shap-ADWIN, a novel approach improving the performance of state-of-the-art drift detectors such as ADWIN by leveraging the information brought by Shapley Loss Values.