Online Next Activity Prediction Under Concept Drifts

Existing research in predictive process maintenance has recently focus on designing models for an online prediction where both training and testing should be performed efficiently, sequentially and in a scalable manner. However, less attention has been given to the realistic requirement of detecting concept drifts on the fly during training. Additionally, the retraining frequency as well as the training dataset size are used to be statically pre-decided. In this work, we address the previous shortcomings when designing our dynamic concept drift detection and online next activity prediction framework DynaTrainCDD. Our framework first uses a PrefixTree-driven method to detect drifts online then utilizes the Weibull distribution to estimate the retraining parameters. An extensive experimental evaluation using 10 real world datasets shows that our model additionally performs on par with or better than state-of-the-art methods in terms of accuracy while requiring a comparable running time.