Enhancing Predictive Process Monitoring with Time-Related Feature Engineering

Predictive process monitoring plays a critical role in process mining by predicting the dynamics of ongoing processes. Recent trends employ deep learning techniques that use event sequences to make highly accurate predictions. However, this focus often overshadows the significant advantages of lightweight, transparent algorithms. This study explores the potential of traditional regression algorithms, namely kNN, SVM, and RF, enhanced by event time feature engineering. We integrate existing and novel time-related features to augment these algorithms and compare their performance against the well-known LSTM network. Our results show that these enhanced lightweight models not only compete with LSTM in terms of predictive accuracy but also excel in scenarios requiring online, real-time decision-making and explanation. Furthermore, despite incorporating additional feature extraction processes, these algorithms maintain superior computational efficiency compared to their deep learning counterparts, making them more viable for time-critical and resource-constrained environments.