Patchlinear for single-step forecasting and anomaly detection

Min-Hua Zheng, Shi-Ying Lan, Wan-Lei Zhao, Jie Zhao

Abstract: The trend forecasting and anomaly detection are the two fundamental tasks on a time series. They become increasingly important due to the popularity of AIOps in various scenarios. In the last few decades, increasingly complicated models have been proposed one after another for the trend forecasting and anomaly detection on the time series. Whereas, they are either ineffective or infeasible as they are too complicated to be deployed in real scenarios, where the online training is barely feasible. In this paper, an effective online method both for single-step forecasting and anomaly detection is proposed. In contrast to the existing methods that employ complicated models, the core component in our method is a two-layer linear network. Although simple, it already outperforms many state-of-the-art methods, such as DLinear and iTransformer on the single-step forecasting task. By the integration of distance significance which detects anomalies by referring to the recent history of the time series, our method also shows considerably superior performance on the anomaly detection task over most of the state-of-the-art methods such as FCVAE and AnomalyTransformer.

External IDs:dblp:journals/mlc/ZhengLZZ25