Skip to contents

Segmentation methods

segmentation-methods.Rmd

This chapter describes the four segmentation classes, the methods they share, and how refitting works. Each class takes a costFunc object and a time series and returns change-points; they differ in the search they use, from greedy (binSeg, Window) to exact (PELT for a penalty, Dynp for a number of change-points).

Classes

After initialising a costFunc object (see Cost functions), create a segmentation object such as binSeg, Window, PELT, or Dynp.

R6 Class Method Description Parameters/active bindings
binSeg Binary Segmentation Recursively splits the signal at points that minimise the cost. minSize, jump, costFunc, tsMat, covariates
Window Slicing Window Detects change-points using local gains over sliding windows. minSize, jump, radius, costFunc,tsMat, covariates
PELT Pruned Exact Linear Time Optimal segmentation with pruning for linear-time performance (costFunc must be PELT-compatible). minSize, jump, costFunc, tsMat, covariates
Dynp Exact Dynamic Programming Globally optimal segmentation for a specified number of change-points, via a full dynamic-programming table rather than binSeg’s greedy search. minSize, jump, nBkpsMax, costFunc, tsMat, covariates

The covariates argument is optional and only required for models involving both dependent and independent variables (e.g., "LinearL2", "LinearSIGMA", "LinearL1"). If not provided, the model is force-fitted using only an intercept term (i.e., a column of ones).

A PELT object, for example, can be initialised as follows:

detectionObj = PELT$new(minSize = 1L, jump = 1L, costFunc = costFunc$new("L2"))

The case studies run these classes from data to change-points: binSeg in Change in mean and variance and Change in autoregressive dynamics, and PELT in Tuning the penalty and Custom cost functions.

Methods

All segmentation objects (binSeg, Window, PELT, Dynp) implement the following methods:

  • $describe(printConfig): Views the (current) configurations of the object.
  • $fit(tsMat, covariates): Constructs a C++ detection module corresponding to the current configurations.
  • $predict(pen, nBkps): Performs change-point detection given a linear penalty value, or a target number of change-points via nBkps (which takes precedence over pen when both are supplied).
  • $eval(a,b): Evaluates the cost of a segment (a,b].
  • $segments(): Returns the cost and parameter estimates of each segment from the latest $predict() (see Segment costs and parameters).
  • $plot(d, endPts,...): Plots change-point segmentation in ggplot style.

binSeg, Window, and Dynp additionally implement:

  • $getHistory(): Returns a data.frame of the cost after 0, 1, 2, ... change-points; for binSeg/Window, also which breakpoint was added at each step (see Model selection for why Dynp’s version omits that column).
  • $plotElbow(maxK): Plots $getHistory()’s cost trajectory against the number of change-points, for choosing nBkps via the “elbow method” instead of tuning pen directly.

Dynp additionally implements $costPath(), the raw numeric vector of exact minimal costs that $getHistory() wraps into a data.frame.

Active bindings and refitting

Active bindings (such as minSize or tsMat) can be modified at any time, before or after the object is created, via the $ operator. For consistency, if the object has already been fitted, modifying any active bindings will automatically trigger the re-fitting process.

detectionObj$minSize = 2L #Before fitting
detectionObj$fit(a_time_series_matrix) #Fitted
detectionObj$minSize = 1L #After fitting - automatically trigger `$fit()`