摘要：The ability to proactively monitor business pro-cesses is a main competitive differentiator for firms. Processexecution logs generated by process aware informationsystems help to make process specific predictions forenabling a proactive situational awareness. The goal of theproposed approach is to predict the next process event fromthe completed activities of the running process instance,based on the execution log data from previously completedprocess instances. By predicting process events, companiescan initiate timely interventions to address undesired devi-ations from the desired workflow. The paper proposes amulti-stage deep learning approach that formulates the nextevent prediction problem as a classification problem. Fol-lowing a feature pre-processing stage with n-grams andfeature hashing, a deep learning model consisting of anunsupervised pre-training component with stacked autoen-coders and a supervised fine-tuning component is applied.Experiments on a variety of business process log datasetsshow that the multi-stage deep learning approach providespromising results. The study also compared the results toexisting deep recurrent neural networks and conventionalclassification approaches. Furthermore, the paper addressesthe identification of suitable hyperparameters for the pro-posed approach, and the handling of the imbalanced nature ofbusiness process event datasets.
关键词：Process prediction; Deep learning; Featurehashing; N-grams; Stacked autoencoders
