Predictive Process Monitoring (PPM) enables organizations to predict future event sequences, numerical indicators, and business outcomes for ongoing process cases. Most existing PPM approaches rely on offline machine learning or deep learning models trained on static event logs. However, these approaches are not well suited to event stream settings, where events arrive continuously, predictions must be produced in real time, and process behavior may evolve due to concept drift. Their practical applicability may also be constrained by prediction latency, computational costs, and the need for repeated full retraining. Streaming methods address these limitations by incrementally updating predictive models as new events arrive. Despite their potential, PPM techniques for event streams have not yet been studied within a unified experimental setting. Existing approaches are evaluated using different pipelines, preprocessing strategies, prediction tasks, and evaluation procedures, making their results difficult to compare and leaving unclear how different adaptive architectures perform on continuously evolving event streams. To address this gap, this thesis designs and implements a standardized end-to-end PPM pipeline specifically for event streams. The pipeline supports the continuous processing of incoming events, the incremental construction of case representations, real-time prediction, and online model adaptation. Using this framework, the thesis conducts a comparative study of different PPM techniques in event stream settings. This work investigates the contribution of individual components within a state-of-the-art streaming deep learning architecture and systematically compares the extensions against classical streaming machine learning models. The approaches are evaluated on multiple benchmark event streams and across the next-activity prediction task. Statistical tests are used to assess the significance of the observed differences between architectural choices.
Predictive Process Monitoring (PPM) enables organizations to predict future event sequences, numerical indicators, and business outcomes for ongoing process cases. Most existing PPM approaches rely on offline machine learning or deep learning models trained on static event logs. However, these approaches are not well suited to event stream settings, where events arrive continuously, predictions must be produced in real time, and process behavior may evolve due to concept drift. Their practical applicability may also be constrained by prediction latency, computational costs, and the need for repeated full retraining. Streaming methods address these limitations by incrementally updating predictive models as new events arrive. Despite their potential, PPM techniques for event streams have not yet been studied within a unified experimental setting. Existing approaches are evaluated using different pipelines, preprocessing strategies, prediction tasks, and evaluation procedures, making their results difficult to compare and leaving unclear how different adaptive architectures perform on continuously evolving event streams. To address this gap, this thesis designs and implements a standardized end-to-end PPM pipeline specifically for event streams. The pipeline supports the continuous processing of incoming events, the incremental construction of case representations, real-time prediction, and online model adaptation. Using this framework, the thesis conducts a comparative study of different PPM techniques in event stream settings. This work investigates the contribution of individual components within a state-of-the-art streaming deep learning architecture and systematically compares the extensions against classical streaming machine learning models. The approaches are evaluated on multiple benchmark event streams and across the next-activity prediction task. Statistical tests are used to assess the significance of the observed differences between architectural choices.
A Comparative Study of Adaptive Predictive Process Monitoring Techniques on Event Streams
SAVINOV, PAVEL
2025/2026
Abstract
Predictive Process Monitoring (PPM) enables organizations to predict future event sequences, numerical indicators, and business outcomes for ongoing process cases. Most existing PPM approaches rely on offline machine learning or deep learning models trained on static event logs. However, these approaches are not well suited to event stream settings, where events arrive continuously, predictions must be produced in real time, and process behavior may evolve due to concept drift. Their practical applicability may also be constrained by prediction latency, computational costs, and the need for repeated full retraining. Streaming methods address these limitations by incrementally updating predictive models as new events arrive. Despite their potential, PPM techniques for event streams have not yet been studied within a unified experimental setting. Existing approaches are evaluated using different pipelines, preprocessing strategies, prediction tasks, and evaluation procedures, making their results difficult to compare and leaving unclear how different adaptive architectures perform on continuously evolving event streams. To address this gap, this thesis designs and implements a standardized end-to-end PPM pipeline specifically for event streams. The pipeline supports the continuous processing of incoming events, the incremental construction of case representations, real-time prediction, and online model adaptation. Using this framework, the thesis conducts a comparative study of different PPM techniques in event stream settings. This work investigates the contribution of individual components within a state-of-the-art streaming deep learning architecture and systematically compares the extensions against classical streaming machine learning models. The approaches are evaluated on multiple benchmark event streams and across the next-activity prediction task. Statistical tests are used to assess the significance of the observed differences between architectural choices.| File | Dimensione | Formato | |
|---|---|---|---|
|
Pavel Savinov. Master Thesis.pdf
accesso aperto
Dimensione
905.61 kB
Formato
Adobe PDF
|
905.61 kB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/115849