Two-Step Anomaly Detection for Time Series Data
Ryan E. Sperl, Soon Myoung Chung · 2019
We propose a new two-step algorithm for anomaly detection in time series data. The first step uses a simple data forecasting model which is fast with high recall but potentially low precision, whereas the second step uses a complex data forecasting model which is slower but more accurate. The first model initially labels each data value. When an anomaly is reported by the first model, the second model relabels that data value, verifying or correcting the initial label. This verification step reduces the false positive rate of the first step. This scheme provides accurate identification of anomalous data values as well as fast rejection of non-anomalous data values. Our experimental results demonstrate higher accuracy compared to the individual component models, albeit with more processing time.