Model-Agnostic Event Log Augmentation for Predictive Process Monitoring

Predictive process monitoring aims to predict how the execution of a running process instance will evolve until its completion. Deep learning techniques have been shown to perform well for various prediction tasks, such as next activity prediction, remaining time prediction, or outcome prediction. However, the quality and performance of these models is highly dependent on the available amount of training data, as deep learning models require a lot of data to generalize well. In practice, the available event logs usually contain only a few thousand records with more or less redundancy, which is insufficient with respect to the large number of parameters that need to be estimated during training. For this reason, data augmentation is often used in machine learning research to increase the amount of available training data by applying transformations to them and create new samples synthetically. Since data augmentation is still largely unexplored in predictive process monitoring, this paper proposes an initial set of simple noise-based transformations that could be applied to any event log and boosts the performance of existing predictive process monitoring approaches. Our experimental evaluation shows that predictive process monitoring approaches for predicting the next activity benefit from this data augmentation technique in terms of performance and stability of the training process.