Explainable Process Deviance Discovery with Data-Efficient Deep Learning

Detecting deviant execution instances of a business process is an important issue in modern enterprises and organizations. Recently, learning deep Deviance Detection Models (DDMs) out of process traces via (semi-)supervised methods was shown to outperform traditional approaches based on standard Machine Learning solutions. However, deep models need to be trained with large enough amounts of data, which may not be available in several real-life settings, especially in Green-AI applications where restrictions to data access and data processing operations must be obeyed [1]. To better suit such settings, a novel approach for discovering a deep DDM is proposed that compensates for the scarcity of training data by leveraging a complementary self-supervised learning task. Besides minimizing the deviance classification loss, the reconstruction loss of an auxiliary auto-encoder sub-net is exploited as an additional (self-)supervision signal. For the sake of explanation, the neural network is provided with a flat pattern-based encoding, on top of which two parallel stacks of feature-representation layers are learnt efficiently and robustly by leveraging residual-like skip connections. The approach effectively deals with the challenging combination of data-efficiency and explainability requirements on a case study concerning the execution traces of a real-life process.