LinearSegmentation

This is a small package that performs linear segmented regression: fitting piecewise linear functions to data, and simultaneously estimating the best breakpoints. Three algorithm are implemented, sliding_window, top_down, and shortest_path.

using LinearSegmentation

segments = segmentation_function(
    x_values, 
    y_values; 
    min_segment_length = minimum_segment_x_length, 
    fit_threshold = minimum_r2,
    fit_function = :r2,
    overlap = true,
)

Where segments = [idxs1, idxs2, ...] is an array of indices, with idxs1 corresponding to the indices of xs in the first segment, idxs2 the second segment, etc. Minimum segment lengths are specified with min_segment_length. By default, the goodness-of-fit is measured using the coefficient of determination (R²). Each segment must have a minimum R² of fit_threshold. Root mean squared error can also be used by setting fit_function = :rmse, and adjusting fit_threshold to a dataset dependent error threshold. In this case, the root mean squared error must be smaller than fit_threshold for each segment. By default, the end of a segment is also the start of the next segment, but this can be changed by setting overlap to false (resulting in disjoint segmentations).

N = 100
xs = collect(range(0, 3 * pi, length = N)) .+ 0.1 .* randn(N)
ys = sin.(xs) .+ 0.1 .* randn(N)

Uses a sliding window approach to segment the data: initially an empty segment is made, and data added to it until fit_threshold is reached. Then a new segment is made, and the process repeats until the data is exhausted. This algorithm is the cheapest to run, but may generate worse fits due to its simplicity.

segments = sliding_window(xs, ys; min_segment_length=1.2)

This algorithm recursively splits the data into two parts, attempting to find segments that are both long enough, and have a good enough fit (set via the kwargs).

segments = top_down(xs, ys; min_segment_length=1.2)

This algorithm is my take on the dynamic programming approaches used by the R packages listed below (NB: not equivalent implementations!). In essence, a weighted directional graph is constructed, where each node corresponds to an index of xs, and the edge weight between nodes corresponds to the goodness-of-fit measure between the two nodes (segment length restrictions and maximum error are both incorporated). The shortest weighted path that spans xs is the found with Graphs.a_star (see Graphs.jl), and should correspond to the best segmentation.

segments = shortest_path(xs, ys; min_segment_length=1.2)

1. https://cran.r-project.org/web/packages/dpseg/vignettes/dpseg.html
2. https://winvector.github.io/RcppDynProg/
3. E. Keogh, S. Chu, D. Hart and M. Pazzani, "An online algorithm for segmenting time series," Proceedings 2001 IEEE International Conference on Data Mining, San Jose, CA, USA, 2001, pp. 289-296, doi: 10.1109/ICDM.2001.989531.