Neural Network Based Drift Detection

Abstract

The unprecedented growth in machine learning has shed light on its unique set of challenges. One such challenge is apparent changes in the input data distribution over time known as Concept Drifts. In such cases, the model’s performance degrades according to the changes in the data distribution. The remedy for concept drifts is retraining the model with the most recent data to improve the model’s performance. The significant issue is identifying the precise point at which the model must be updated for maximum performance benefits with minimum retraining effort. This problem is challenging to address in unsupervised detection methods with no access to label data to identify the changing distributions for the targets of the input data. Here, we present our unsupervised method based on a Generative Adversarial Network and a feed forward neural network for detecting concept drifts without the need for target labels. We demonstrate that our method is better at identifying concept drifts and outperforms the baseline and other comparable methods.