What is it about?
Predictive process monitoring uses data from previous process executions to anticipate what may happen next in a process that is still running. However, many existing approaches assume that business processes remain stable over time. In real-world environments, process behaviour can change, making models trained on past data progressively less reliable. This paper introduces ATLAS, an adaptive approach for predicting the next activity of an ongoing process. ATLAS combines deep learning with online learning: instead of remaining fixed after its initial training, the predictive model is continuously updated whenever a new event is observed. The approach uses a pretrained language model to represent process activities and an LSTM neural network to learn their sequential behaviour. It is designed to deal with evolving data and concept drift while keeping the computational requirements of continuous model updating under control.
Featured Image
Photo by National Cancer Institute on Unsplash
Why is it important?
Organizations increasingly rely on process predictions to anticipate delays, undesirable events and operational problems. These predictions are only useful, however, when the underlying model reflects the process as it currently operates. ATLAS moves predictive process monitoring from a static learning setting towards a continuously adaptive one. By learning from incoming events, it can respond to changes in process behaviour without requiring the predictive system to be rebuilt manually from scratch. The study evaluates this approach in terms of both predictive accuracy and computation time and compares it with existing online predictive process-monitoring methods. This joint consideration is important because operational systems need models that are not only accurate, but also efficient enough to be updated as data arrive.
Perspectives
Our research explores predictive process monitoring as a streaming learning problem rather than as a conventional offline prediction task. Business processes evolve, and predictive models should evolve with them. With ATLAS, we investigate how deep neural models can learn continuously from process events while supporting next-activity prediction for cases that are still running. The broader goal is to develop predictive process-monitoring systems that remain useful and responsive in dynamic organizational environments. Key takeaways Business processes may change over time because of concept drift. Predictive models trained only on historical data can become outdated. ATLAS predicts the next activity of ongoing process executions. The model is updated continuously as new process events arrive. The method combines online learning, language-model-based representations and an LSTM network. The evaluation considers both predictive accuracy and computational efficiency.
Prof. Donato Malerba
Universita degli Studi di Bari Aldo Moro
Read the Original
This page is a summary of: Predictive Process Monitoring Through the Lens of Deep Online Learning, July 2026, Springer Science + Business Media,
DOI: 10.1007/978-3-032-32643-0_20.
You can read the full text:
Contributors
The following have contributed to this page
