gSeg-package: Graph-Based Change-Point Detection

In gSeg: Graph-Based Change-Point Detection (g-Segmentation)

Description

This package can be used to estimate change-points in a sequence of observations, where the observation can be a vector or a data object, e.g., a network. A similarity graph is required. It can be a minimum spanning tree, a minimum distance pairing, a nearest neighbor graph, or a graph based on domain knowledge.

For sequence with no repeated observations, if you believe the sequence has at most one change point, the function gseg1 should be used; if you believe an interval of the sequence has a changed distribution, the function gseg2 should be used. If you feel the sequence has multiple change-points, you can use gseg1 and gseg2 multiple times. See gseg1 and gseg2 for the details of these two function.

If you believe the sequence has repeated observations, the function gseg1_discrete should be used for single change-point. For a changed interval of the sequence, the function gseg2_discrete should be used. The function nnl can be used to construct the nearest neighbor link.

Author(s)

Hao Chen, Nancy R. Zhang, Lynna Chu, and Hoseung Song

Maintainer: Hao Chen ([e-mail])

References

Chen, Hao, and Nancy Zhang. (2015). Graph-based change-point detection. The Annals of Statistics, 43(1), 139-176.

Chu, Lynna, and Hao Chen. (2019). Asymptotic distribution-free change-point detection for modern data. The Annals of Statistics, 47(1), 382-414.

Song, Hoseung, and Hao Chen (2020). Asymptotic distribution-free change-point detection for data with repeated observations. arXiv:2006.10305

See Also

gseg1, gseg2, gseg1_discrete, gseg2_discrete, nnl

Examples

data(Example)
# Five examples, each example is a n-length sequence.
# Ei (i=1,...,5): an edge matrix representing a similarity graph constructed on the
# observations in the ith sequence.  
# The following code shows how the Ei's were constructed.

require(ade4) 
# For illustration, we use 'mstree' in this package to construct the similarity graph.
# You can use other ways to construct the graph.

## Sequence 1: change in mean in the middle of the sequence.
d = 50
mu = 2
tau = 100
n = 200
set.seed(500)
y = rbind(matrix(rnorm(d*tau),tau), matrix(rnorm(d*(n-tau),mu/sqrt(d)), n-tau))
y.dist = dist(y)
E1 = mstree(y.dist)
# For illustration, we constructed the minimum spanning tree.
# You can use other ways to construct the graph.

r1 = gseg1(n,E1, statistics="all")  
# output results based on all four statistics
# the scan statistics can be found in r1$scanZ

r1_a = gseg1(n,E1, statistics="w")  
# output results based on the weighted edge-count statistic

r1_b = gseg1(n,E1, statistics=c("w","g"))  
# output results based on the weighted edge-count statistic 
# and generalized edge-count statistic

# image(as.matrix(y.dist))  
# run this if you would like to have some idea on the pairwise distance

## Sequence 2: change in mean away from the middle of the sequence.
d = 50
mu = 2
tau = 45
n = 200
set.seed(500)
y = rbind(matrix(rnorm(d*tau),tau), matrix(rnorm(d*(n-tau),mu/sqrt(d)), n-tau))
y.dist = dist(y)
E2 = mstree(y.dist)
r2 = gseg1(n,E2,statistic="all")
# image(as.matrix(y.dist))  


## Sequence 3: change in both mean and variance away from the middle of the sequence.
d = 50
mu = 2
sigma=0.7
tau = 145
n = 200
set.seed(500)
y = rbind(matrix(rnorm(d*tau),tau), matrix(rnorm(d*(n-tau),mu/sqrt(d),sigma), n-tau))
y.dist = dist(y)
E3 = mstree(y.dist)
r3=gseg1(n,E3,statistic="all")
# image(as.matrix(y.dist)) 


## Sequence 4: change in both mean and variance away from the middle of the sequence.
d = 50
mu = 2
sigma=1.2
tau = 50
n = 200
set.seed(500)
y = rbind(matrix(rnorm(d*tau),tau), matrix(rnorm(d*(n-tau),mu/sqrt(d),sigma), n-tau))
y.dist = dist(y)
E4 = mstree(y.dist)
r4=gseg1(n,E4,statistic="all")
# image(as.matrix(y.dist))  


## Sequence 5: change in both mean and variance happens on an interval.
d = 50
mu = 2
sigma=0.5
tau1 = 155
tau2 = 185
n = 200
set.seed(500)
y1 = matrix(rnorm(d*tau1),tau1)
y2 = matrix(rnorm(d*(tau2-tau1),mu/sqrt(d),sigma), tau2-tau1)
y3 = matrix(rnorm(d*(n-tau2)), n-tau2)
y = rbind(y1, y2, y3)
y.dist = dist(y)
E5 = mstree(y.dist)
r5=gseg2(n,E5,statistics="all")
# image(as.matrix(y.dist))  

## Sequence 6: change in mean away from the middle of the sequence 
## when data has repeated observations.
d = 50
mu = 2
tau = 100
n = 200
set.seed(500)
y1_temp = matrix(rnorm(d*tau),tau)
sam1 = sample(1:tau, replace = TRUE)
y1 = y1_temp[sam1,] 
y2_temp = matrix(rnorm(d*(n-tau),mu/sqrt(d)), n-tau)
sam2 = sample(1:tau, replace = TRUE)
y2 = y2_temp[sam2,] 
y = rbind(y1, y2)
# Data y has repeated observations
y_uni = unique(y)
E6 = nnl(dist(y_uni), 1)
cha = do.call(paste, as.data.frame(y))    
id = match(cha, unique(cha))
r6 = gseg1_discrete(n, E6, id, statistics="all")
# image(as.matrix(y.dist))  

gSeg documentation built on Oct. 23, 2020, 5:54 p.m.