DTW Barycenter Averaging
A global averaging method for time series under DTW (Petitjean, Ketterlin and Gancarski 2011).
DBA( X, centroid = NULL, ..., window.size = NULL, norm = "L1", max.iter = 20L, delta = 0.001, error.check = TRUE, trace = FALSE, mv.ver = "by-variable" ) dba( X, centroid = NULL, ..., window.size = NULL, norm = "L1", max.iter = 20L, delta = 0.001, error.check = TRUE, trace = FALSE, mv.ver = "by-variable" )
X: A matrix or data frame where each row is a time series, or a list where each element is a time series. Multivariate series should be provided as a list of matrices where time spans the rows and the variables span the columns of each matrix.
centroid: Optionally, a time series to use as reference. Defaults to a random series of X
if NULL. For multivariate series, this should be a matrix with the same characteristics as the matrices in X.
...: Further arguments for dtw_basic(). However, the following are already pre- specified: window.size, norm (passed along), and backtrack.
window.size: Window constraint for the DTW calculations. NULL means no constraint. A slanted band is used.
norm: Norm for the local cost matrix of DTW. Either "L1" for Manhattan distance or "L2" for Euclidean distance.
max.iter: Maximum number of iterations allowed.
delta: At iteration i, if all(abs(centroid_{i} - centroid_{i-1}) < delta), convergence is assumed.
error.check: Logical indicating whether the function should try to detect inconsistencies and give more informative errors messages. Also used internally to avoid repeating checks.
trace: If TRUE, the current iteration is printed to output.
mv.ver: Multivariate version to use. See below.
The average time series.
This function tries to find the optimum average series between a group of time series in DTW space. Refer to the cited article for specific details on the algorithm.
If a given series reference is provided in centroid, the algorithm should always converge to the same result provided the elements of X keep the same values, although their order may change.
The windowing constraint uses a centered window. The calculations expect a value in window.size that represents the distance between the point considered and one of the edges of the window. Therefore, if, for example, window.size = 10, the warping for an observation considers the points between and , resulting in 10(2) + 1 = 21 observations falling within the window.
The indices of the DTW alignment are obtained by calling dtw_basic() with backtrack = TRUE.
Please note that running tasks in parallel does not guarantee faster computations. The overhead introduced is sometimes too large, and it's better to run tasks sequentially.
This function uses the RcppParallel package for parallelization. It uses all available threads by default (see RcppParallel::defaultNumThreads()), but this can be changed by the user with RcppParallel::setThreadOptions().
An exception to the above is when it is called within a foreach parallel loop made by dtwclust . If the parallel workers do not have the number of threads explicitly specified, this function will default to 1 thread per worker. See the parallelization vignette for more information - browseVignettes("dtwclust")
This function appears to be very sensitive to numerical inaccuracies if multi-threading is used in a 32 bit installation. In such systems, consider limiting calculations to 1 thread.
There are currently 2 versions of DBA implemented for multivariate series (see examples):
mv.ver = "by-variable", then each variable of each series in X and centroid are extracted, and the univariate version of the algorithm is applied to each set of variables, binding the results by column. Therefore, the DTW backtracking is different for each variable.mv.ver = "by-series", then all variables are considered at the same time, so the DTW backtracking is computed based on each multivariate series as a whole. This version was implemented in version 4.0.0 of dtwclust, and it is faster, but not necessarily more correct.# Sample data data(uciCT) # Obtain an average for the first 5 time series dtw_avg <- DBA(CharTraj[1:5], CharTraj[[1]], trace = TRUE) # Plot matplot(do.call(cbind, CharTraj[1:5]), type = "l") points(dtw_avg) # Change the provided order dtw_avg2 <- DBA(CharTraj[5:1], CharTraj[[1]], trace = TRUE) # Same result? all.equal(dtw_avg, dtw_avg2) ## Not run: # ==================================================================================== # Multivariate versions # ==================================================================================== # sample centroid reference cent <- CharTrajMV[[3L]] # sample series x <- CharTrajMV[[1L]] # sample set of series X <- CharTrajMV[1L:5L] # the by-series version does something like this for each series and the centroid alignment <- dtw_basic(x, cent, backtrack = TRUE) # alignment$index1 and alginment$index2 indicate how to map x to cent (row-wise) # the by-variable version treats each variable separately alignment1 <- dtw_basic(x[,1L], cent[,1L], backtrack = TRUE) alignment2 <- dtw_basic(x[,2L], cent[,2L], backtrack = TRUE) alignment3 <- dtw_basic(x[,3L], cent[,3L], backtrack = TRUE) # effectively doing: X1 <- lapply(X, function(x) { x[,1L] }) X2 <- lapply(X, function(x) { x[,2L] }) X3 <- lapply(X, function(x) { x[,3L] }) dba1 <- dba(X1, cent[,1L]) dba2 <- dba(X2, cent[,2L]) dba3 <- dba(X3, cent[,3L]) new_cent <- cbind(dba1, dba2, dba3) # sanity check newer_cent <- dba(X, cent, mv.ver = "by-variable") all.equal(newer_cent, new_cent, check.attributes = FALSE) # ignore names ## End(Not run)
Petitjean F, Ketterlin A and Gancarski P (2011). ``A global averaging method for dynamic time warping, with applications to clustering.'' Pattern Recognition, 44 (3), pp. 678 - 693. ISSN 0031-3203, tools:::Rd_expr_doi("10.1016/j.patcog.2010.09.013") , https://www.sciencedirect.com/science/article/pii/S003132031000453X.