Demo
In the age of digital transformation, the efficient management of business operations is highly dependent on the Business Process Management (BPM) solutions employed. This type of software solution allows businesses to automate monotonous tasks; design and manage their operational workflows. As BPM software lies at the heart of a business’s operational management, given they are underdeveloped or primitive, they can be highly detrimental to a business’s lifecycle.
Given their nature, this type of software generates vast amounts of data in the form of process event logs. These logs are digital footprints of actions and events taking place in the real world. Given that, any deviation or mishap (anomaly) taking place in the real world will be reflected on these logs. With the aim of detecting said deviations, our foundation started the project AnomalyNet, an anomaly detection framework for BPM solutions supported by the Scientific and Technological Research Council of Turkey (TUBITAK).
Our Published Paper
Our conducted research with the title Multi-Graph Anomaly Detection in Business Processes With Scalable Neural Architectures was published in IEEE Access journal, which is a Science Citation Index Expanded (SCI-e) journal, an index which only the journals with the highest quality and impact criteria indexed in. It has a quality index of Q1 and an H index of 247. Given its high impact factor and wide reach, IEEE Access is one of the most respected open access journals in the software and engineering world and is a valuable platform for researchers.
Our Research
Our study focuses on the usage of a multi-graph structure and graph autoencoders (GAE) for business process anomaly detection. By blending state-of-the-art feature engineering and deep learning approaches, our research team successfully closed the existing gaps and overcame the weaknesses of existing approaches. As a result, a deep learning model which will act as the brain of the AnomalyNet framework was developed.
The realized deep learning model was tested on both public and private datasets and has shown up to 20% higher performance in anomaly detection in comparison to other state-of-the-art approaches. Simultaneously, thanks to the utilization of transformer architecture, up to 60% training and inference time was saved alongside higher detection performance. These results achieved on injections based on real-world anomaly scenarios suggest that the model is fit for real-world applications and will be a valuable tool for business process optimization.
The architectural diagram of the proposed decoder improvements. Above, transformer-only decoder is represented. Below, the transformer and gated recurrent unit (GRU) hybrid decoder is represented.
If you wish to learn more about our research, we invite you to read our research paper below:
