Explanations of Anomalies on Graph Structured Business Processes

Within the context of business process management (BPM) and process mining, anomalies, often defined as deviations from the standard flow of a given process, has the potential to significantly impact a businesses lifecycle. Anomalies which can represent anything from a simple inefficiency to fraudulent activity, is often times reflected within the event logs of the digitalized business processes. The detection of said deviances can help boost a businesses efficiency and protect it against fraudulent activity. Furthermore, by understanding the causes of these deviances businesses can further optimize their processes increasing their profitability. In this study we leveraged two different approaches for explainable artificial intelligence on graph structured business process data. Initially, we build and test a total of 8 different Graph AutoEncoder (GAE) models consisting of 4 unique encoders and 2 different decoders as the key components of the architecture. After that we test dimension reducing standard AutoEncoder (AE) models with a novel type of AE working towards increasing the dimensionality of the latent feature representations. Once all the testing is done we pick the champion model and use the labels acquired from it to train a predictive Graph Neural Network (GNN) model to get features importances by leveraging two different types of explainable artificial intelligence approaches meant to work on graph structured data: GNNExplainer and GraphLIME. Using the aforementioned methodology, we analyze and unravel the anomalies within the logs of an anonymous tenant of the Next4biz BPM platform.