Local-to-Global Unsupervised Anomaly Detection from Temporal Data

Anomaly detection for temporal data has received much attention by many real-world applications. Most existing unsupervised methods dealing with this task are based on a sequential two-way approach (clustering and detection). Because of this, the clustering is less robust to anomalous series in data which distorts the detection step. Thus, to overcome this problem, we propose an embedded technique simultaneously dealing with both methods. We reformulate the task of anomaly detection as a local-weighting-instance clustering problem. The anomalous series are detected locally in each cluster as well as globally in the data, as a whole. Extensive experiments on benchmark datasets are carried out to validate our approach and compare it with other state-of-the-art methods of detection.