Process models are widely used to describe operational processes, e.g., producing a car, treating a patient, or handling an insurance claim. The ML4ProM project introduces predictive Machine Learning technology that allows detecting early-on possible problems or deviations from the original model, which can be avoided or counteracted using predictive analytics.

Process Mining deals with discovering a process (lower left part), e.g. detecting how patients are treated and handled in a hospital. In ML4ProM Machine Learning will be used to detect deviations from that process through predictions and propose how to adapt to changes of reality. Illustration: Christoph J Kellner, Studio Animanova

Project Overview

In many areas, like logistics, industrial production, treating a patient in a hospital (as shown as an example in the illustration above) or finance, event data are used to describe and compile symbolic process models. As such process models can be learned, they provide a formal and interpretable description of the underlying events (see illustration: learning in step (1), (2), (3)). Often process mining technologies are retrospectively applied in order to validate an existing model or identify problematic parts of a model. The ML4Pro aims at complementing such retrospective analysis through predictive Machine Learning technology in order to detect deviation from the model as early as possible and propose possible remedies (right part in illustration).

The approach will combine process models–that describe event knowledge in a symbolic way, for example, using Petri-nets–and Machine Learning techniques–as recurrent neural networks that will be tasked to learn drifts of critical parts of (sub-)processes. Process models are easy explainable and can be used to communicate predictions and proposed actions. Hence, it is important that explainability is maintained in the extension towards learned components, for example, using relevance learning which identifies relevant features in relation to a process.




    1. 1.
      Artelt A, Hammer B. Efficient computation of counterfactual explanations of LVQ models. Published online August 2019.
    2. 2.
      Losing V, Hammer B, Wersing H. Self-Adjusting Memory: How to Deal with Diverse Drift Types. In: ; 2017:4899-4903. doi:10.24963/ijcai.2017/690