CHANGELOG.md

In asardaes/dtwclust: Time Series Clustering Along with Optimizations for the Dynamic Time Warping Distance

Changelog

Version 5.5.12

• Remove explicit C++ requirements.

Version 5.5.11

• Documentation update (#61).
• Specify minimum supported version of package Matrix.

Version 5.5.10

• Enhanced compare_clusterings to allow complex parameters in lists (#59).

Version 5.5.9

• Removed some deprecated interfaces from the C++ code.

Version 5.5.8

• For compare_clusterings, fixed a bug that caused the order of columns "k" and "method" to be wrong in the data frame with results for hierarchical configurations.

Version 5.5.7

• Series order is now preserved in plots of included shiny apps.
• The optimization in sdtw_cent now uses stats::optim by default.
• Dropped strong dependency on bigmemory in case it is archived. It can still be used if available.

Version 5.5.6

• Adjustments for future R 4.0.0 version: matrices now also inherit from array.

Version 5.5.5

• Function dtw_basic gained parameter sqrt.dist.

Version 5.5.4

• Fixed compilation on Solaris.

Version 5.5.3

• Fixed a bug in the way compare_clusterings handled preprocessing of series.
• Internal speed optimizations.
• Added support for interruption of multi-threaded tasks via RcppThread.
• Updated timing experiments.

Version 5.5.2

• Fixed plotting of clusters with labels when a cluster has repeated series.

Version 5.5.1

• Documentation fixes.
• The registration of hclust and proc_time with setOldClass now leaves them as virtual classes.

Version 5.5.0

• Added a standalone function to extract centroids for partition around medoids: pam_cent.
• Fixed the calculation of GAK for non-NULL windows.
• "Fixed" the calculation of internal CVIs for clusterings done with TADPole.
• Vignette updates.

Version 5.4.1

• Fixed the distance calculation done by the dtw2 wrapper for multivariate series.

Version 5.4.0

• Added another helper for compare_clusterings: cvi_evaluators.
• Fixed the calculation of SF and CH CVIs for hierarchical clustering with default centroid extraction.
• The data frames passed to pick.clus in compare_clusterings now also contain the configurations' data instead of just the scores. This will require adjustment of existing pick.clus functions.
• The centroids in the objects returned by tsclust now have an attribute series_id if said centroids were extracted from the provided series without modification. It is an integer vector indicating which series were chosen as centroids.
• Improved warning messages in compare_clusterings.
• Multi-threaded parallelization is now more medium-grained to hopefully balance load a bit better.
• Minor updates to parallelization vignette.

Version 5.3.1

• Added parameter no.expand to compare_clusterings_configs.
• Documentation fixes.
• You should now be able to use this package without attaching it, with some caveats. See documentation for dtwclust-package.

Version 5.3.0

• Added 2 shiny apps, see interactive_clustering and ssdtwclust.
• Added another helper for compare_clusterings: repeat_clustering.
• Added distmat to fuzzy_control for when centroid = "fcmdd".
• Dropped dependency on rngtools. Changing the random number generator after loading dtwclust should not be a problem now.
• Fixed a bug regarding the way tsclust handled the seed when called from compare_clusterings. This will affect compare_clusterings results of hierarchical or TADPole clusterings made therein if a centroid subject to randomness was used (e.g. dba).

Version 5.2.0

• Several functions now use RcppParallel for parallelization, refer to the new parallelization vignette for more information.
• None of the distance/centroid functions expose parameters for helper matrices anymore.
• Added the option to add labels to non-dendrogram plots.
• The ellipsis in sdtw_cent is now actually passed to nloptr.
• Fixed tsclust for dtw_lb distance and pam.precompute = FALSE (it was always using dtw_basic, so the results should not change).
• Updated timing experiments.
• Added soft-DTW information to the PDF vignette.

Version 5.1.0

• Added implementations of soft-DTW distance and centroid.
• Fixed the calculation of the clusinfo slot for TSClusters class for clusterings that ended with empty clusters.
• Fixed logic in partitional/fuzzy clustering, which was preventing the usage of parallel backends when only one value of k and one repetition were made.

Version 5.0.0

• Removed all deprecated code.
• Added external fuzzy cluster validity indices.
• Fixed GAK's sigma estimation in the non-proxy version. The estimation routine in the proxy version also changed slightly, and the estimates will be different due to randomness.
• Fixed the calculation of the Silhouette CVI (#21).
• Modified tslist to hopefully make it faster.
• Partitional and fuzzy controls (tsclust-controls) now specify version 2 of the main loop by default.
• Packages clue and ggplot2 are no longer automatically attached by default.
• Fixed minor potential bugs in several functions.
• Changed plyr dependency for dplyr.
• Minor vignette updates.

Version 4.1.1

• Added some fuzzy cluster validity indices to cvi.
• When using TADPole clustering through tsclust, custom parameters can now be passed to a custom centroid function if provided.
• Modified the partitional/fuzzy loops, which were doing an extra iteration. Previous version still available for backwards compatibility. This modification was wrong in version 4.1.0.

Version 4.0.3

• Deactivating PAM precomputation will no longer use sparse matrices by default, explicitly set pam.sparse to TRUE if you want this functionality.
• Fixed multivariate plots (#18).
• Fixed zscore for data frame input (it was still coercing column-wise instead of row-wise).
• Added an additional vignette with some timing experiments.

Version 4.0.2

• Ported the proxy versions of lb_keogh, lb_improved, SBD, GAK and dtw_basic to C++ and improved them by using the bigmemory package.
• Ported TADPole to C++.
• Ported part of the algorithm that updates sparse distance matrices to C++.
• Improved the optimizations for symmetric matrices that are calculated in parallel.
• Fixed tsclustFamily's dist function for matrix or data frame input.
• Fixed partitional PAM centroids for dtw_lb distance and pam.precompute = FALSE (#16).
• Exported a function to coerce matrices and data frames to a list.
• Updated documentation.

Version 4.0.1

• Ported dtw_lb to C++ when using dtw_basic.
• Modified some tests to account for rounding error (CRAN request).

Version 4.0.0

• Optimized TADPole for multiple k and dc values.
• Optimized PAM centroids with pam.precompute = FALSE by using sparse matrices from the Matrix package.
• Optimized shape_extraction by using the eigs_sym function from the RSpectra package.
• Implemented the DTW lower bounds in C++.
• Implemented DBA in C++.
• Implemented an alternative version of multivariate DBA that might be faster. See its documentation.
• Added a symmetric control for fuzzy clustering.
• Partitional, hierarchical and fuzzy configurations in compare_clusterings now take into account the symmetric control parameter.
• The functionality for pick.clus in compare_clusterings changed depending on the value of return.objects.
• Fixed an error that sometimes caused objects returned by tsclust to have duplicated elements in the args slot.
• Fixed DTW symmetry detection for fuzzy clustering.
• Some internal functions changed, so older objects might no longer be compatible. Try using update(old_TSClusters_obj).
• The dtwclust function is now deprecated. Try using as(dtwclust_class_obj, "TSClusters") for old objects.
• Changed name of function compute_envelop to compute_envelope (old one still available but deprecated).
• Minor vignette updates.
• Several internal optimizations.

Version 3.2.0

• Added functions compare_clusterings and helpers to compare many clustering configurations, possibly in parallel.
• Fixed an error in tsclust that prevented CVIs to be calculated for hierarchical/TADPole cases if a custom centroid function was used.
• Added slot seed to the objects returned by tsclust.

Version 3.1.2

• The arguments in tsclust's ellipsis are now passed to all preprocessing, centroid and distance functions.
• Fixed some symmetry detection in tsclust when using DTW.
• Updated vignette to use tsclust in the examples.

Version 3.1.1

• Seeds are now set when calling TADPole through dtwclust with a parallel backend and multiple values of k, in case the centroid function has randomness associated.
• Added a new experimental function tsclust that should be functionally equivalent to dtwclust but is hopefully more coherent in general.

Version 3.1.0

• Fixed subsetting of multivariate plots
• Implemented the memory-saving version of DTW in dtw_basic when backtrack = FALSE.
• Implemented fuzzy c-medoids.
• Seeds were not being set in hierarchical/TADPole clustering, which could affect reproducibility of preprocessing/centroid functions.
• The force.symmetry argument in dtw_lb was removed since it didn't serve any real purpose.
• Exported the function to compute envelops.
• Minor vignette updates.

Version 3.0.0

• Bear in mind that the DTW/SBD algorithms (and hence the functions that depend on them) might give different results in installations with 32-bit architectures.
• Removed deprecated arguments/slots.
  • If you have older dtwclust objects saved, try updating them with attr(dtwclust_object, "centers") <- NULL if you run into compatibility problems.
  • DBA arguments changed order.
• Removed the dba.alignment argument from DBA since other step.patterns don't really work.
• Added (conditional) support for more hierarchical procedures. See the examples and vignette.
• Added a new distance based on global alignment kernels: GAK.
• Added support for functions in package clue.
• Fixed detection of some symmetric DTW cases.
• No longer enforcing preprocessing/centroid functions with ellipsis in their formals.
• Added a multivariate dataset sample: CharTrajMV.
• Improved plots for clusterings with multivariate series.
• Updated vignette.

Version 2.3.0

• Correction: DTW can be symmetric for series of both equal/different length, although in general this is not necessarily the case, due to asymmetric step patterns or constrained paths.
• Exported the dtw2 function.
• Revamped zscore and reinterpolate functions.
• Data frames are now parsed row-wise, like matrices, to maintain consistency with proxy.
• Fixed a bug where multivariate series with different length had spurious data added to them.
• Fixed a bug in multivariate shape_extraction. Reminder: multivariate shape extraction might not be a good idea.
• Consistency adjustments: all center(s) arguments/slots will be removed and replaced with centroid(s).
• Added the dtw_basic function, which should be faster due to its limited functionality. It can also be used with dtw_lb and DBA. It will now be used by default.
• The SBD and shape_extraction functions used the crossprod function internally, which returned a 1x1 matrix by default, causing some dimension inconsistencies. In the future, R will give an error about the inconsistency, so the function has been changed in dtwclust, but it resulted in very small numerical differences, which may be enough to alter some clustering results.

Version 2.2.1

• Added package vignette
• Implemented several cluster validity indices in the new cvi function
• The custom argument force.pairwise used in some of the proxy distances is not necessary anymore
• No longer enforcing NULL dimensions for each series (in case multivariate series are to be used), use with caution
• Fixed an error that prevented fuzzy clustering from working when k = 2
• Fixed a possible problem in some of the included proxy distances when a data object had a length of 1

Version 2.1.2

• Added option to specify a function to extract prototypes in hierarchical and TADPole clustering
• Switched algorithm to calculate envelops to Lemire's, which should be slightly faster
• More plot types
• Documentation fixes

Version 2.1.1

• Minor bug fixes for fuzzy clustering
  • Fixed the predict generic.
  • The final values returned in the fcluster slot needed one final update during clustering. It should be correct now, but it will vary slightly with respect to what was previously given in v2.1.0.
• Some more examples

Version 2.1.0

• Added fuzzy c-means clustering
• Memory optimizations for DBA
• Support for several k values for multiple runs
• Update documentation

Version 2.0.0

• Major refactor.
• Many formal parameters from the dtwclust function were dropped and implemented in the formal class dtwclustControl. For now, they will still be supported through ... with a message.
• No longer supporting non-proxy distances, in order to be able to take advantage of the included optimizations.
• Dropped inheritance of flexclust's kccasimple
  • Many slots and methods were ported
  • Inheriting from hclust class now, and all its associated methods
• shape_extraction now accepts series with different lengths!
• More parallel support
  • DBA and Shape centroid calculations
  • DBA itself (probably unnecessary unless you're averaging a lot of series)
  • Custom proxy distances directly (except DTW2)
• Several hierarchical procedures can be made in one run.
• Added distmat slot and update generic to save some time if possible. See examples of dtwclust.
• Extra parameters for distance functions should be correctly detected now.
• Using dtw_lb function now correctly warns about pam.precompute being TRUE.
• Option to calculate pairwise distances with DTW_LB, SBD, LB_Keogh and LB_Improved. See their respective notes. The distance function created for the dtwclustFamily slot also supports this.
• Clusters are randomly re-initialized if they become empty at some iteration.
• Now all included centroid functions recompute centers only if necessary.
• Option to optimize distmat calculation if the distance function is symmetric.

Version 1.3.0

• Added the possibility to run several repetitions for partitional procedures, using different random starts each time by using the doRNG package
• Parallel computing all around. Watch out for RAM! Use Linux if possible.
  • Repetitions can be done in parallel if the user provides a suitable backend
  • Calculation of distance matrices also takes advantage of parallel backends.
  • TADPole and dtw_lb too, probably negligible (maybe even detrimental) for small datasets.

Version 1.2.0

• Added the option to prevent pre-computation of distance matrix when using PAM centroids
• Added an example with a custom distance function
• Using closures instead of relying on passing environments as attributes
• Optimized the SBD function registered with proxy, it's a lot faster now
• Optimized clustering with DBA and shape_extraction so that centers are only recomputed if necessary
• Added processing time slot to class definition
• Several bug fixes (especially in case custom distances were used, and also for plot method)
• Fixed minor bugs in TADPole

Version 1.1.0

• Added more options to the plot method for custom time labels
• Fixed an error in the calculation of Lemire's improved lower bound
• Added L2 norm to DBA, as well as window constraint
• Optimized DBA calculations
• More examples

Version 1.0.0

• Limited support for time series of different lengths
• Improved plot method
• More control parameters
• Added more slots to the dtwclust class
• Some parameters changed order
• Default values of some auxiliary functions were changed (DBA and SBD)
• No longer supporting native kccaFamilies, supporting any distance registered in proxy instead
• Several bug fixes, especially if dtw or dtw2 were being used, in which case they may have been erroneously computed before

Version 0.1.0

• Initial release

asardaes/dtwclust documentation built on March 3, 2023, 5:32 a.m.