Incremental Predictive Process Monitoring: The Next Activity Case

Next-activity prediction methods for business processes are always introduced in a static setting, implying a single training phase followed by the application of the learned model during the test phase. Real-life processes, however, are often dynamic and prone to changes over time. Therefore, all state-of-the-art methods need regular retraining on new data to be kept up to date. It is, however, not straightforward to determine when to retrain nor what data to use; for instance, should all historic data be included or only new data? Updating models that still perform at an acceptable level wastes a potentially large amount of computational resources while postponing an update too much will deteriorate model performance. In this paper, we present incremental learning strategies for updating these existing models that do not require fully retraining them, hence reducing the number of computational resources needed while still maintaining a more consistent and correct view of the process in its current form. We introduce a basic neural network method consisting of a single dense layer. This architecture makes it easier to perform fast updates to the model and enables us to perform more experiments. We investigate the differences between our proposed incremental approaches. Experiments performed with a prototype on real-life data show that these update strategies are a promising way forward to further increase the power and usability of state-of-the-art methods.