A survey of methods for time series change point detection
Citations
506 citations
120 citations
80 citations
Cites background or methods from "A survey of methods for time series..."
..., Aminikhanghahi and Cook (2017) or Truong et al. (2020). Note that we focus here on the unsupervised change point detection problem, without access to external data that can be used to tune hyperparameters (Hocking et al., 2013a; Truong et al., 2017). Let yt ∈ Y denote the observations for time steps t = 1, 2, . . ., where the domain Y is d-dimensional and typically assumed to be a subset of Rd. A segment of the series from t = a, a + 1, . . . , b will be written as ya:b. The ordered set of change points is denoted by T = {τ0, τ1, . . . , τn}, with τ0 = 1 for notational convenience. In offline change point detection we further use T to denote the length of the series and define τn+1 = T + 1. Note that we assume that a change point marks the first observation of a new segment. Early work on CPD originates from the quality control literature. Page (1954) introduced the CUSUM method that detects where the corrected cumulative sum of observations exceeds a threshold value....
[...]
..., Aminikhanghahi and Cook (2017) or Truong et al. (2020). Note that we focus here on the unsupervised change point detection problem, without access to external data that can be used to tune hyperparameters (Hocking et al....
[...]
..., Aminikhanghahi and Cook (2017) or Truong et al. (2020). Note that we focus here on the unsupervised change point detection problem, without access to external data that can be used to tune hyperparameters (Hocking et al., 2013a; Truong et al., 2017). Let yt ∈ Y denote the observations for time steps t = 1, 2, . . ., where the domain Y is d-dimensional and typically assumed to be a subset of Rd. A segment of the series from t = a, a + 1, . . . , b will be written as ya:b. The ordered set of change points is denoted by T = {τ0, τ1, . . . , τn}, with τ0 = 1 for notational convenience. In offline change point detection we further use T to denote the length of the series and define τn+1 = T + 1. Note that we assume that a change point marks the first observation of a new segment. Early work on CPD originates from the quality control literature. Page (1954) introduced the CUSUM method that detects where the corrected cumulative sum of observations exceeds a threshold value. Theoretical analysis of this method was subsequently provided by Lorden (1971). Hinkley (1970) generalized this approach to testing for differences in the maximum likelihood, i....
[...]
...An alternative view of evaluating CPD algorithms considers change point detection as a classification problem between the “change point” and “non-change point” classes (Killick et al., 2012; Aminikhanghahi and Cook, 2017)....
[...]
...For a more expansive review of change point detection algorithms, see e.g., Aminikhanghahi and Cook (2017) or Truong et al. (2020)....
[...]
74 citations
74 citations
Cites background from "A survey of methods for time series..."
...A good representation of the literature on this problem is surveyed in detail in a recent paper in [1]....
[...]
...For clarity, this latter task is sometimes called “semantic segmentation” [1][31]; where there is no danger of confusion, and we will refer to it as segmentation....
[...]
...In essence, many researchers overlay the results of the segmentation on the original data, and we invite the reader to confirm that it matches human intuition [1][5][17][19]....
[...]
...While there are many techniques for segmentation [1][15] [17][19][25], they all have one or more limitations that have prevented their utilization in real world settings....
[...]
...• Most research efforts in this domain test on limited datasets [1][15]....
[...]
References
3,830 citations
"A survey of methods for time series..." refers background in this paper
...The most familiar change point algorithm is cumulative sum [7,10,18,33], which accumulates deviations relative to a specified target of incomingmeasurements and indicates that a change point exists when the cumulative sum exceeds a specified threshold....
[...]
1,812 citations
"A survey of methods for time series..." refers background in this paper
...Definition 2 A stationary time series is a finite variance process whose statistical properties are all constant over time [57]....
[...]
1,750 citations
"A survey of methods for time series..." refers background in this paper
...Here, the observation at each time point is the digital encoding of an image [47]....
[...]
1,693 citations
1,320 citations
"A survey of methods for time series..." refers methods in this paper
...This graph can be defined based on a minimum spanning tree [26], minimum distance pairing [52], nearest neighbor graph [26,52], or visibility graph [40,66]....
[...]