A Contrastive Transformer-Enhanced Temporal Hybrid Deep Learning Framework for Robust Workflow Log Anomaly Detection

Abstract

Detecting anomalies in workflow and event logs is an extremely important component in achieving process reliability, security and efficiency in insurance claim management systems. Conventional machine learning methods tend to fail to realize long-range correlations and complicated deviations in sequences of procedures. This paper introduces a hybrid deep learning system, consisting of Transformer, Convolutional Neural Network (CNN), Temporal Convolutional Network (TCN), Variational Autoencoder (VAE), and contrastive learning to achieve high-quality log-level anomaly detection. Transformers, CNN and TCN extract global, local and temporal features respectively. VAE learns normal process behaviour by probabilistically reconstituting, and contrastive learning learns more effectively in the presence of a class imbalance. Hybrid anomaly scoring algorithm is a scoring method that is used to improve the robustness of reconstruction errors and classification probabilities. The TPE-based Bayesian optimization is used to optimize hyperparameters in terms of F1-score. Empirical findings on an insurance workflow log show that it has a better performance, with an accuracy of 97.8 per cent and F1-score of 0.97 and is better than LSTM, GRU, and standalone Transformer baselines. The proposed framework offers a powerful and scalable algorithm of detecting workflow anomalies.