Segment costs and parameters

This chapter shows how to get the cost and parameter estimates of each segment, either from a fitted segmentation ($segments()) or for any segment you choose, without running a detection algorithm (costFactory).

library(rupturesRcpp)

Costs and parameters of each segment

After $predict(), $segments() returns a list with one element per segment (available on binSeg, Window, PELT and Dynp; the example below uses PELT). Each element is a list with Start, End, Cost and Params, where the segment is (Start, End] with the same 0-based convention as $eval(a, b). Cost equals $eval(Start, End), and Params is the same named list costFactory’s $get_params() returns (see below).

set.seed(1)
tsMat = cbind(c(rnorm(100, 0), rnorm(100, 5, 5)))

PELTObj = PELT$new(minSize = 5L, costFunc = costFunc$new("SIGMA"))
PELTObj$fit(tsMat)
PELTObj$predict(pen = 50)
#> [1] 100 200

segs = PELTObj$segments()
segs[[2]]
#> $Start
#> [1] 100
#> 
#> $End
#> [1] 200
#> 
#> $Cost
#> [1] 312.2758
#> 
#> $Params
#> $Params$mean
#> [1] 4.81096
#> 
#> $Params$cov
#>          [,1]
#> [1,] 22.70893

For PELT, the segment costs add up to the optimal penalised cost minus pen times the number of change-points.

sapply(segs, `[[`, "Cost")
#> [1] -22.47755 312.27581

Re-fitting (including through an active binding such as $minSize or $costFunc) clears the segmentation saved by the last $predict(), so $predict() must be run again before calling $segments().

Costs without a detection algorithm

Sometimes you don’t need a detection algorithm at all, only fast cost evaluation and parameter estimation for segments whose boundaries you already have (e.g., cross-validating a pen value against known change-points, or just querying a segment’s fitted parameters). costFactory wraps the same C++ cost modules used internally by PELT/binSeg/Window/Dynp, exposing only $eval(), $get_params() and $segments(), with no segmentation logic.

set.seed(1)
tsMat = cbind(c(rnorm(100, 0), rnorm(100, 5, 5)))

cf = costFactory$new(costFunc$new("L2"))
cf$fit(tsMat)
cf$eval(0, 100)
#> [1] 79.86945

cf$get_params(0, 100)
#> $mean
#> [1] 0.1088874

As with the segmentation classes, $costFunc is an active binding: reassigning it after $fit() automatically re-fits the underlying module against the same data.

cf$costFunc = costFunc$new("SIGMA")
#> `costFunc` has been updated. Re-fitting the model.

cf$eval(0, 100)
#> [1] -22.47755

cf$get_params(0, 100)
#> $mean
#> [1] 0.1088874
#> 
#> $cov
#>           [,1]
#> [1,] 0.7986955

$get_params()’s return shape depends on the cost function: median/mean for "L1"/"L2", mean and cov for "SIGMA", coef (intercept first) for "VAR"/"LinearL2"/"LinearL1", coef and cov for "LinearSIGMA", and params (whatever paramFun returns) for "Custom". Both covs are biased estimates, divided by the segment length n with no bias or degrees-of-freedom correction, and include epsilon on the diagonal if addSmallDiag = TRUE.

$segments(endPts) runs $eval() and $get_params() over every segment of a given segmentation, returning the same list as the segmentation classes’ $segments() (see above). endPts must end at n, so the output of any $predict() can be passed as is:

segs = cf$segments(c(100, 200))
sapply(segs, `[[`, "Cost")
#> [1] -22.47755 312.27581