Enhancing Business Process Anomaly Detection with Sequence Capturing Hybrid Autoencoders

Since the introduction of the original deep network architecture, Multi-Layer Perceptron (MLP), the research in the deep learning domain took of with researchers proposing various deep network architectures each with their respective strengths and shortcomings. One method to combine the strength of these neural networks it to combine them, creating what is known as a hybrid model, which can be an effective way to construct deep networks for domains where machine learning (ML) is applicable. One such domain, where specialized neural networks can yield considerable performance increase is the Business Process Anomaly Detection domain. Our study explores the usage of sequence capturing autoencoders to extract featurerich latent feature representations working in tandem with convolutional neural networks (CNN), for additional feature extraction to enrich the latent features extracted from business process sequences. Our results show that combining different architectures can yield notable performance increases up to $8 \%$ in event level anomaly detection (from 0.552 to 0.633 in F1-score) and 3% in trace level anomaly detection (from 0.734 to 0.766 in F1-score), with additional empirical findings related to hybrid deep model construction.