piwik-script

Intern
    Wirtschaftsinformatik

    Artikel im Journal "Decision Support Systems" angenommen

    11.01.2021

    Die Autoren zeigen, dass Gated Convolutional Neural Network (GCNN) und das Key-Value-Predict Attention Network (KVP) bei variierender Sparsity, Variation und Repetitiveness sehr gute Ergebnisse für das Predictive Process Monitoring liefern können.

    Das von Kai Heinrich, Patrick Zschech, Markus Bonin und Prof. Janiesch verfasste Research Paper "Process Data Properties Matter: Introducing Gated Convolutional Neural Networks (GCNN) and Key-Value-Predict Attention Networks (KVP) for Next Event Prediction with Deep Learning" untersucht die Eignung verschiedener maschineller Lernverfahren gegeben variierender Prozesseigenschaften. Eine erste Fassung des Papiers basierend auf denselben Daten war auf der 15. internationale Tagung Wirtschaftsinformatik 2020 in Postdam erschienen und hatte dort den Best Paper Award gewonnen.

    Decision Support Systems ist eines der wichtigsten Outlets der internationalen Wissenschafts-Community für Themen der technologiegestützten Entscheidungsfindung. Die VHB listet es als B.

    Abstract:

    Predicting next events in predictive process monitoring enables companies to manage and control processes at an early stage and reduce their action distance. In recent years, approaches have steadily moved from classical statistical methods towards the application of deep neural network architectures, which outperform the former and enable analysis without explicit knowledge of the underlying process model. While the focus of prior research was on the long short-term memory network architecture, more deep learning architectures offer promising extensions that have proven useful for other applications of sequential data. In our work, we introduce a gated convolutional neural network and a key-value-predict attention network to the task of next event prediction. In a comprehensive evaluation study on 11 real-life benchmark datasets, we show that these two novel architectures surpass prior work in 34 out of 44 metric-dataset combinations. For our evaluation, we consider the effects of process data properties, such as sparsity, variation, and repetitiveness, and discuss their impact on the prediction quality of the different deep learning architectures. Similarly, we evaluate their classification properties in terms of generalization and handling class imbalance. Our results provide guidance for researchers and practitioners alike on how to select, validate, and comprehensively benchmark (novel) predictive process monitoring models. In particular, we highlight the importance of sufficiently diverse process data properties in event logs and the comprehensive reporting of multiple performance indicators to achieve meaningful results. https://doi.org/10.1016/j.dss.2021.113494

    Die Implementierung der Verfahren und weitere Forschungsdaten finden sich unter https://doi.org/10.23728/b2share.08b7ff704f724b94a61b4a6cac0fe1e0.

    Zurück