Benchmarking the xLSTM Architecture for Business Process Anomaly Detection

Business process anomaly detection helps organizations uncover deviations in their processes, which often indicate severe inefficiencies or malicious activity. The article by Teoman Berkay Ayaz, Alper Özcan and Akhan Akbulut puts a new type of recurrent neural network to the test in this domain: xLSTM (Extended Long-Short Term Memory).

The study uses xLSTM to build unsupervised autoencoders for business process anomaly detection and compares it with three mainstream architectures: LSTM, GRU and Transformer. To keep the comparison fair, every model is built with an identical structure and the same hyperparameters; the models differ only in their core recurrent or self-attention blocks. Each model is also evaluated with and without bidirectionality and, where applicable, multi-head attention.

According to the empirical results, the xLSTM-based autoencoder was the most consistent and robust architecture across seven datasets, reaching the highest mean F1-scores: 0.476 at trace level and 0.258 at event level.

The article was published on June 29, 2026 in the Kocaeli Journal of Science and Engineering (2026, number 17, pp. 93-103), a journal of Kocaeli University. The full text is open access on the journal's DergiPark page. Our other work on business process anomaly detection includes SPECTRE and AnomalyNet.