Wraps changepoints::WBS.network() (Yu, Padilla, Wang and Rinaldo):
wild binary segmentation on a sequence of networks, detecting the times at
which the edge-probability structure changes. The input is one row per
time point holding the vectorised adjacency matrix, so a series of
\(p \times p\) networks over \(n\) times is an \(n \times p^2\)
matrix.
Usage
network_wrapper(
x,
copy2 = NULL,
n_intervals = 100,
threshold = NULL,
alpha = 0.05,
n_perm = 20,
delta = NULL,
seed = NULL
)Arguments
- x
The network sequence: an \(n \times p^2\) matrix of vectorised adjacency matrices, or an \(n \times p \times p\) array.
- copy2
An independent second observation of the same network sequence, in the same shape. The method's guarantees rest on sample splitting; see the section below for what happens when there is only one copy.
- n_intervals
Number of random intervals. Defaults to
100.- threshold
Detection threshold. When
NULL, calibrated by permutation exactly as inhdcov_wrapper().- alpha, n_perm
Level and number of permutations for that calibration.
- delta
Minimum spacing. Defaults to
max(5, floor(n / 20)).- seed
Optional seed.
When you have only one copy of the network
WBS.network() takes two independent observations of the sequence,
which is how the theory controls the bias of the squared-Frobenius
statistic. Given a single sequence of binary networks this wrapper
constructs the second copy by splitting each edge indicator at random
(each present edge is assigned to one copy with probability one half),
which is the usual independent-thinning device and is reported in a
message. For weighted networks it splits the weight instead, which is
exact for Poisson weights and approximate otherwise. If you have a genuine
replicate, pass it as copy2 and none of this applies.
References
Yu Y, Padilla OHM, Wang D, Rinaldo A (2021). “Optimal network online change point localisation.” arXiv preprint arXiv:2101.05477. doi:10.48550/arXiv.2101.05477 .
See also
Other changepoint engines:
bcp_wrapper(),
beast_wrapper(),
bfast_wrapper(),
binsegrcpp_wrapper(),
bocpd_wrapper(),
cpm_wrapper(),
cpop_wrapper(),
cpt_wrapper(),
decafs_wrapper(),
ecp_wrapper(),
envcpt_wrapper(),
esac_wrapper(),
fabisearch_wrapper(),
fastcpd_wrapper(),
fcov_wrapper(),
fmean_wrapper(),
fpop_wrapper(),
geomcp_wrapper(),
hdcov_wrapper(),
hdreg_wrapper(),
idetect_wrapper(),
inspect_wrapper(),
kcp_wrapper(),
kwc_wrapper(),
mcp_wrapper(),
mosum_wrapper(),
not_wrapper(),
npmojo_wrapper(),
nsp_wrapper(),
ocd_wrapper(),
pilliat_wrapper(),
segmented_wrapper(),
smuce_wrapper(),
sn_wrapper(),
strucchange_wrapper(),
taylor_wrapper(),
tguh_wrapper(),
trend_wrapper(),
var_wrapper(),
wbs2_wrapper(),
wbs_wrapper(),
wbsts_wrapper()
Examples
set.seed(2026)
p <- 5
mk <- function(n, prob) {
t(replicate(n, as.numeric(matrix(stats::rbinom(p * p, 1, prob), p))))
}
X <- rbind(mk(40, 0.2), mk(40, 0.6))
network_wrapper(X, n_intervals = 20, n_perm = 20, seed = 1)
#> No independent second observation supplied: splitting each edge at random to build one. See the "When you have only one copy" section of ?network_wrapper.
#> ggcpt (changepoint detection result)
#> Method: network
#> Change in: network
#> Changepoints found: 1
#> CP convention: left
#> Penalty: threshold = 3.2294
#> Series length: 80
#>
#> Changepoints:
#> # A tibble: 1 × 2
#> cp cp_value
#> <int> <dbl>
#> 1 40 0.08