Efficient state change detection method for modeling stateful time series data

Abstract

AbstractTime series analysis is an important research area and has wide real applications. Time series modeling has been identified and well known as the key fundamental problem in time series analysis, and it has received lots of research attentions. The hardness of achieving sophisticated and efficient time series models depends on the characteristic of the data met. In practical, time series data are usually generated by the physical process with several working states, which are determined by the system design. In such cases, the data characteristics of each state is quite different from the other states, where the modeling problem becomes much harder and it is more challenging to design modeling methods for stateful time series data. Therefore, this paper focuses on the stateful time series data, and studies the state change detection problem, which is identified as a key fundamental problem for modeling stateful data. To achieve effective and efficient state change detection, the proposed method utilizes two essential ideas. Since the discrepancy (a.k.a. histogram) based method has been proposed and verified to be more effective by previous works, the first idea is to improve its performance by involving more temporal information and utilizing the encoder–decoder techniques to reform the data. In view of the efficiency issue of the discrepancy‐based method, the second idea is to integrate both the Bayesian and discrepancy‐based detection methods, such that a perfect tradeoff between time efficiency and detection performance can be achieved. Experiments on both real and synthetic datasets show that the proposed methods are both effective and efficient, and the state change detection problem can be solved well by them on stateful time series data.