Predictive Process Monitoring (PPM) is a discipline within business process mining that leverages historical process event logs to predict the outcome of ongoing process instances. This capability enables organizations to transition toward proactive decision-making, optimizing their operations. However, the practical application of PPM is significantly hindered by the inherent noisiness of real-world process event logs. Such noise can result from deviations from standard business process definitions, infrequent process activities or paths, data recording errors, or an imbalance in the representation of process case variants in the log. This could negatively impact the performance and stability of machine learning models used in PPM, as atypical entries can prevent the models from effectively learning the underlying process, leading to poor predictive accuracy and limited generalization when deployed in live environments. This paper investigates the impact of event log noise on PPM performance. A PPM machine learning algorithm is trained on multiple, increasingly filtered versions of a real-world business process event log. Each version represents an attempt to systematically exclude noisy event log entries to achieve different levels of denoising. The experimental results quantitatively demonstrate the significant influence of event log data quality on the efficacy of machine learning models, showing that targeted denoising of the event log training data directly enhances the model's predictive performance and robustness. This study underscores the importance of pre-processing and noise mitigation as a crucial step in developing reliable and generalizable PPM solutions.

Keywords

Process Denoising, Predictive Process Monitoring, Business Process Mining, and Machine Learning.