Unified model for collective and point anomaly detection using stacked temporal convolution networks

摘要

Time-series anomaly detection utilizing deep learning methods is widely used in fraud detection, network intrusion detection, and medical anomaly detection. Most deep learning methods exclusively focus on models based on recurrent neural networks (RNNs), such as long short-term memory (LSTM) or gated recurring units (GRUs), rather than on models based on convolutional neural networks (CNNs) or integrated ones. Inspired by the success of CNN-based models in many scenarios, we propose a single model that can be used for detecting both collective and point anomalies using stacked temporal convolution networks (CPA-TCN). Compared with state-of-the-art models, the CPA-TCN model boasts the following advantages. First, the CPA-TCN model reconstructs sequential features with current inputs and historical features and is only trained on normal datasets. Second, the CPA-TCN model outperforms RNN-based models in terms of speed and accuracy across diverse tasks and datasets and demonstrates more effective memory. Third, the CPA-TCN model can effectively detect both collective and point anomalies by detecting point anomalies before collective anomalies, overcoming the shortcomings of models that can either detect point or collective anomalies. Fourth, the two-part anomaly detection module can significantly improve the accuracy of point anomaly detection. Extensive experiments on real-world datasets demonstrate that our CPA-TCN model achieves better prediction results with the ROC-AUC of 98%–99% compared to state-of-the-art methods and thus has a competitive advantage.