NEWS

Version 0.4.1

  2026-07-18

1. Version 0.4.1 created from 0.4 (or 0.4.0).
2. In VII_clustering.cpp, moved int current_threads = 1; into #ifdef _OPENMP
3. Revised the second example of discrete.jointly(): replaced z=tan(x) by z=x^2. The original code creates a degenerate case when k > 2.
4. Added documentation for parameters of discretize.jointly(), ballclust(), and cluster() functions.
5. Added documentation for hashed_upsilon() function.

Version 0.4.0

  2026-07-02

1. Improved: Finalized the package vignettes and added comprehensive tests.

  2026-06-19

1. Added: Implemented a custom, high-performance spherical Gaussian Mixture Model (VII covariance structure) backend, bypassing previous computational overheads.
2. Added: Introduced automated K-selection driven by BIC, utilizing k-means for fast centroid initialization.
3. Added: Integrated OpenMP multithreading support to the E-step of the clustering algorithm, providing near-linear speedups on multi-core systems.
4. Changed: Replaced the dependency on ‘mclust’ with the new VII model, significantly improving the scalability of the core algorithm and unlocking efficient processing for high-dimensional datasets.

Version 0.3.2.1

  2026-06-04

1. Added automatic naming of returned grid based on the input variable names.

Version 0.3.2

  2025-12-11

1. Created version 0.3.2 from 0.3.1.2.
2. Revised DESCRIPTION.
3. Updated documentation for the exported gen_simdata() function.
4. Updated testthat tests in test-Ball-approx_likelihood.R
5. Updated vignette Examples.Rmd
6. Deleted the Makevars file, resulting in replacing fixed CXX14 to default CXX17 C++ compiler.
7. Cleaned up warning messages in C++ code due to signed / unsigned mismatch
8. Commented out the #pragma STDC FENV_ACCESS ON in Joint_Grid.cpp.
9. Moved ’static const vector *spx_c;’ from Cluster.h to Cluster.cpp.
10. Updated documentation of function discretize.jointly()
11. Commented out some options in discretize.jointly() to the cluster_method argument. They are methods not yet ready for release.

Version 0.3.1.2

  2025-12-08

1. [Bug fix] When cutting in decreasing order, the cutting positions (lines) were also sorted in decreasing order, causing only the first line to be used when constructing the table. The lines are now reordered in increasing order to prevent this issue.

  2025-12-02

1. Added: Implemented bidirectional cutting—attempt cuts in both increasing and decreasing order, then select the best cut.
2. Added: Introduced a new option for the grid_method parameter in the R interface, enabling users to apply a bidirectional evaluation strategy that tests cuts in both increasing and decreasing order and selects the best result.

Version 0.3.1.1

  2025-11-28

1. Added: Introduced a new return value dimensional_cll that stores the best BIC scores across all dimensional cuts.

Version 0.3.1

  2025-11-04

1. The function now uses std::thread::hardware_concurrency() to detect available CPU cores on all supported platforms (including Linux).
2. Added automatic validation to handle invalid or excessive thread numbers and ensure at least one thread is used even when system detection fails.
3. Corresponding documentation for thread parameter (nthread) updated.

  2025-11-07

1. improved numerical consistency: adjusted C++ backend to ensure identical double-precision results across all platforms (macOS ARM, macOS Intel, Linux, and Windows).

Version 0.3.0.3

  2025-10-23

1. Minor issue fixed in c++. (boundary incorrect for dp-binomial version)

  2025-8-21

1. Minor issue fixed in c++. (boundary incorrect for dp-recursion version)

Version 0.3.0.2

  2025-2-06

1. Added robustness to the C++ implementation, ensuring it correctly handles data sets where all data points belong to a single cluster (i.e., all labels are identical).”

  2025-2-04

1. Introduced a new evaluation metric, mid_dist, which computes the Euclidean distance between points and the median of their respective blocks.
2. Optimized the C code to enhance reliability, improve execution speed, and increase readability.

Version 0.3.0.1

  2024-7-20

1. Fixed a bug in the C++ implementation related to purity calculation.
2. Added additional evaluation scores to the return value.
3. Included discretization labels in the return value.
4. Improved the plot function to display column names as x-axis or y-axis labels, if available.

Version 0.3.0

  2024-4-23

1. added “telescope” as a clustering method

  2024-4-15

1. Added dqRNGkind(“xoroshiro128+”) before dqset.seed(…) in testthat files to avoid crash by the new version of dqrng

  2024-3-16

1. Added a checking if mclust result is NULL which could lead the code crash when the data is very small and dense as mclust cannot working on such data. If mclust result id NULL, we will add some jitter noise to the data to let the code working correctily, which makes the code more robust.

Version 0.2.0

  2023-10-03

1. Implements multithreads in c++ to speedup the process. User are allowed to input the number of thread they want to use, if it exceed the max number of thread that machine has, it will automatically set to the max number of thread provided by function hardware_concurrency().
2. Update the test file as many parameters has been added.
3. Fixed some bugs.

Version 0.1.7

  2023-09-02

1. Added a new parameter “scale”, the default value is false. If the data variance on each dimension differ to much, you are strongly suggested set it to true to obtain a better clustering result.Please noted the scaled data will only be used in the process of clustering.

Version 0.1.6.1

  2022-09-13

1. Added “const” keyword to parameters of some functions.
2. Added some type_cast to improve the robust.
3. Some miner structure improvement and optimization.

Version 0.1.6

  2022-08-27

1. Added a new parameter “entropy”, the default value is false. If user set it to true, then the score for the dynamic programming matrix will be calculated based on entropy instead of likelihood.
2. entropy score: N(D_x) * log(N(D_x)/N(D)) likelihood score: N(D_x) * N(D)* log(N(D_x)/N(D))

Version 0.1.5

  2022-08-25

1. Made minor changes to eliminate C/C++ warning messages in BIC.h, BIC.cpp, Cutting_Cluster_dp.cpp, Cutting_Cluster_dp_compressed.cpp, Joint_Grid.h, Joint_Grid.cpp.

2022-08-20

To check:

1. In Example 9, ruby’s dataset, when grid_method = “Sort+Split”, it will cut at 0, however, no line should be put if the left side equal to the right side. Check C code.

2022-08-17

1. Rename the grid_method “Dp compressed likelihood” to “DP exact likelihood”, “Dp compressed likelihood nlog” to “Dp approx likelihood”.

2. Removed ULONG_MAX in “Cut_by_density.cpp” to fix compiling issues under Linux. Similarly, DBL_MAX and INT_MIN were removed in Cutting_Cluster_dp_compressed.cpp.

3. “#include ” was added to “Joint_Grid.cpp” to fix a compiling error under Linux.

4. Removing the usage of “cref()” in “Cutting_Cluster_dp.cpp” and “Cutting_Cluster_dp_compressed.cpp” to avoid compiling error on Linux.

5. Adding test files for “DP Compressed likelihood nlogn” and “DP Compressed likelihood”in testthat.

6. Adding the perimeter “noise” for user to determined whether adding jitter noise to original data or not. The default value is FALSE which no noise will be added. Only will be used when cluster_method = “BALL+BIC”.

7. Adding a new grid method “DP Compressed likelihood nlogn”, which is a speed up version of “DP Compressed likelihood” by using Divide and Conquer when filling the matrix, but no optimal will be generated. The “DP Compressed likelihood nlogn” is set to default for the parameter “grid_method”.

8. Optimizing the structure of the cluster class in c code to speed up and make the code more robust, and rewrite the “Sort+Split” part correspondingly.

9. Adding cluster_method and grid_method to the result, if the cluster label is given by user, the cluster_method will be set to “user supply”.

10. Modify the plot function to print both methods in title.

Version 0.1.4

2022-08-07

1. Replaced ULONG_MAX by std::numeric_limits::max() in “Cut_by_density.cpp” to fix compiling issues under Linux. Header file is included. Similarly, DBL_MAX and INT_MIN were replaced in Cutting_Cluster_dp_compressed.cpp.

2. Included header file to fix compiling error on Linux regarding cref() in “Cutting_Cluster_dp.cpp” and “Cutting_Cluster_dp_compressed.cpp”.

3. “#include ” was added to “Joint_Grid.cpp” to fix a compiling error under Linux.

2022-07-27

4. Created from version 0.1.3.

Version 0.1.3

To-do:

1. To fix. The Examples.Rmd vignette does not compile on MacOS.

2. To fix. Five test cases fail.

2022-06-30

1. Created from version 0.1.2.

2. Used dynamic programming to maximize the likelihood of categorical distribution, to improve quality of discretization

3. Used BIC on the categorical likelihood to select number of discrete levels for each variable.

4. Compression: avoid trying discretization within consecutive points of the same cluster label to greatly reduce runtime of dynamic programming.

Version 0.1.2

2022-03-17

1. Added an additional parameter cutoff, default value is 0.05. This parameter is only used for the “DP” and “DP Compressed”. After calculate the line for each dimension, we will calculate the p-value. If the p-value is larger than the cutoff, no line will be kept for that dimension. Otherwise, store those lines.

Version 0.1.1

2022-03-10

1. Added new cluster methods “DP” and “DP Compressed”. The new method using dynamic programming to find the optimal solution to split the label into many zone. The maximum number of zone is set to 2*K by hand currently.

2. Implemented a new approach of the “DP” method. Improved by compressed the label, the size of the dynamic programming table can be decreased dramatically.

3. Before return the bins on each dimension, we will check the upsilon static to make sure the cutting is reasonable. Currently the cut off has been set to the median of the PDF. Any result with a upsilon static that smaller than the median of the PDF will be drop and no cut will be put on that dimension.

Version 0.1.0

2022-01-25

1. When cluster_method = “Ball+BIC”, if Mclust discretization was unsuccessful, discretize.jointly will now add jitter noise (of increasing ‘amount’) to ‘data’.

2022-01-18

1. Updated CITATION, REFERENDCES.bib, and README.md.

2022-01-17

1. Fixed a bug in computing weights for MultiChannel.WUC when cluster IDs are not consecutive numbers after Ball clustering by Mclust (descrtize_jointly.R:198-204)
2. Fixed function cluster(data, k, method) where the method was not included into the return object.
3. Examples in the vignettes have been updated.

2022-01-11

1. Added k=Inf option, that incrementally increases the number of clusters to choose until either an increment is not required or each data point is its own cluster.
2. Added an example showing the use of k=Inf option.

2021-10-26

1. Added an additional parameter ‘cluster_method’, could be “Ball+BIC”, “kmeans+silhouette” or “PAM”,
2. Added clustering method Mclust
3. Changed the default cluster method from Kmeans to Mclust, which is “Ball+BIC”
4. Added an additional parameter ‘grid_method’, could be “Sort+split” or “MultiChannel.WUC”, “MultiChannel.WUC” is a discretize method Multichannel.MUC from Ckmeans.1d.dp. The default option is still “Sort+split” which is the same as previous version. As the “MultiChannel.WUC” option is experimental, it is not recommended.
5. Added general plotting function plot(), old plot function still available
6. Changed the default value of min_level from 2 to 1. We expanded min_level argument to provide a vector of integers specifying the minimum level for each dimension.

Version 0.0.8.1

2021-10-06

1. Removed the requirement for the number of points in the data

Version 0.0.8

2020-09-13

1. Added function plotGOCpatterns to plot the continuous data along with the cluster preserving grid.
2. Created a manual for the plotGOCpatterns() function.
3. Updated the code for the Examples vignette to use plotGOCpatterns.

2020-08-10

1. Created version 0.0.8 from 0.0.7.
2. Added an additional parameter ‘min_level’ to denote the minimum number of discretization levels required for each dimension.
3. Updated the manual of discrete.jointly() function.
4. Added an entry in reference and citation.
5. Updated README.md with badges.

Version 0.0.7

2020-04-03

1. Tidied up code for the Examples vignette.
2. Updated the manual of discrete.jointly() function.
3. Made minor editorial changes in DESCRIPTION and README.md.
4. Resolved signed/unsigned mismatches.

2020-03-31

1. Created version 0.0.7 from 0.0.6.
2. Fixed memory leak in Clusters.cpp when calculating median.

Version 0.0.6

2020-03-26

1. Created version 0.0.6 from 0.0.5.
2. Rewrote Prep_Index() to work in between two consecutive points, rather than on top of a single point.
3. Using distance() in Prep_Index() to calculate the distance for two iterators.
4. Using “ceil” in Binary_Index_Searching() to consider even/odd cases when determining grid lines.
5. Fixed potential overflow issues.

Version 0.0.5 (not released to the public)

2020-03-25

1. Fixed a bug in the prep_index() function.
2. Fixed prep_index() (lines 120 and 125) such that grid lines are put at the midpoint between two conseuctive points, instead of on one of the points.
3. Updated vignette. Example 2 seems always correct now.

2020-03-24

1. Created version 0.0.5 from 0.0.4.
2. Function discretize.jointly() now returns cluster labels of each observation and a similarity score (ARI) between the joint discretization and the cluster labels of each observation.
3. The class Cluster has a new constructor that takes cluster labels and the input data to compute median for each dimension.
4. Find_grid() is now based on median.
5. Using ‘dqrng’ in test cases to avoid RNG issue in testing.
6. Rewrote multiple functions in Joint_Grid.cpp to avoid push_back().
7. New visualization code in vignette now shows cluster labels for each observation.

Version 0.0.4 (not released to the public)

2020-03-20

1. Created version 0.0.4 from 0.0.3.
2. Fixed typos in DESCRIPTION and README files.

Version 0.0.3

2020-03-17

1. Created version 0.0.3 from 0.0.2. Package renamed to GridOnClusters
2. Function joint.grid.discretize.R() renamed to discretize.jointly()
3. Return values of function discretize.jointly() changed to include both the discretized data and the grid
4. Manual for discretize.jointly() updated.
5. Line 104, 105 in Joint_Grid.cpp commented out
6. Rewrote Find_Grid() to avoid push_back() in Joint_Grid.cpp
7. Created a vignette to include examples.

Version 0.0.2 (not released to the public)

2020-03-14

1. Created the initial version 0.0.1. Package renamed to QNJGD

Version 0.0.1 (not released to the public)

2020-03-09

1. Created the initial version 0.0.1. Package named JointGridDiscr