Data-Efficient Deep Learning Approach for Explainable Process Deviance Discovery

Detecting deviant execution instances (e.g., related to security attacks, frauds, and faults~\cite{Dumas14}) of a business process is relevant for modern enterprises and organizations. Recent works~\cite{ismis22} demonstrated that learning deep \emph{Deviance Detection Models} (DDMs) out of process traces via (semi-)supervised methods outperforms traditional approaches based on standard Machine Learning solutions. However, deep models require to be trained with a large number of examples, which may not be available in real-life settings, especially in Green-AI applications where restrictions to data access and processing operations could be enforced~\cite{adadi2021}. To better suit such challenging applications, we propose a novel approach to discovering a deep DDM that deals with the scarcity of training data by leveraging a complementary self-supervised learning task. %\cite{mtl}. Specifically, the DDM discovery problem is formalized as a Multi-Task Learning one that simultaneously minimizes the classification loss and the reconstruction error of an auxiliary auto-encoder sub-net. This articulated learning scheme allows us to grasp an additional source of supervision complementary to data labels. For the sake of interpretability and faithful explanation of predictions, our DDM is trained with a tabular representation of process traces~\cite{Dumas14}, on top of which two parallel stacks of feature-representation layers are learned efficiently (and robustly to overfitting) by leveraging residual-like skip connections. The approach effectively deals with the challenging combination of data efficiency and explainability requirements in a case study concerning the execution traces of a real-life process. Future research will investigate alternative multi-objective optimization strategies and scalable uncertainty quantification methods to identify out-of-distribution instances.