Change Log

All notable changes to this project will be documented in this file. This change log follows the conventions of keepachangelog.com. For releases before version [1.0.0] this project did not follow semver.
For releases after version 1.0.0 this project will follow semver, with the addition that minor releases might only add additional tests and/or documentation updates rather than bug fixes.

[13.0.4] - 2023-12-26

This releases gets working support for GPU backed emd calculations up to <256 categories. It supports dataset sizes that other k means implementations I've tried fail to support. It is also faster than most other k means implementations I've tried in part because it uses a better algorithm than k-means++.

However, if you aren't on the happy path of EMD and k-means-afk this release probably breaks everything.

[1.0.0] - 2022-12-04

This release improves the usability of loading and saving clustered models. It is potentially backward incompatible with previous releases because it changes the returned output of the cluster result record.

Added

• ClusterResult now provides an :assignments field which is a filepath to an assignments dataset.
• ClusterResult now provides a .load-centroids method.

Changed

• ClusterResult now has a filepath rather than a tech.ml.dataset under the :centroids key.
• ClusterResult :configuration field now contains a map rather than the full clustering state record.
  This map contains the following configuration fields: :k, :m, :distance-key, :init and :col-names.
• ClusterResult .load-assignments now returns a sequence of sequences.
• The ClusterResult classify method now supports dynamic dispatch. When given a vector it will assume that vector has fields in :col-name ordering. When given a map, it will extract fields in :col-name ordering before calling classify.

Fixed

• The load-model function would fail to load persisted cluster results from disk because when printed to a file the output of a dataset would have pretty printing which conflicted with the reader syntax.
• Calling .load-assignments on a massive cluster result should no longer triggers out of memory issues.

Testing

• Unit testing added for .save-model.
• Unit testing added for load-model.
• Unit testing added for .classify.
• Unit testing added for .load-assignments.
• Unit testing added for .load-centroids.
• Unit testing added for the ClusterResult :configuration field validity.

[0.3.4] - 2022-12-03

Fixed

• Expands unit test coverage.
• Fixes a bug which could cause keywords in datasets to trigger clustering failures.

[0.3.4] - 2022-12-03

Fixed

• Fixes a bug which could cause incorrect assignment of points to clusters.
• Expands unit test coverage.

[0.3.1] - 2022-11-28

Added

• Added support for saving and loading learned models.
• Assignments when calling k-means-seq no longer overwrite each other.
• Add .load-assignments and .classify to ClusterResult.

Changed

• Updated documentation to reflect usage.
• Improved ease of use - defaults for chain length chosen automatically when not provided.
• Expanded test coverage with generative testing.

[0.1.5] - 2022-06-08

Added

• Now support the mc^2 initialization method.
• Now support the afk mc^2 initialization method.

[0.1.4] - 2022-05-28

Added

• k-means can now be called with an option to specify which distance function to use.
• k-means now supports many different distance functions.

Changed

• Default calling convention now runs multiple instances of k means clustering.

[0.1.3] - 2022-05-25

Added

• Now supporting initialization via k-means++.
• Now supporting initialization via k-means||.
• Now supporting initialization via k-means||.
• Now supporting initialization via naive uniform sampling.
• k-means can now be called with options to determine the initialization method.
• Now supporting parquet format.
• Now supporting arrow format.
• Now supporting arrows format.
• Added logs to help make progress of computations more obvious.
• Added the initialize-centroids multimethod; choice of initialization method can now be controlled by callers via multimethods.
• Now support calling k-means with lazyseqs which are ->dataset compatible.
• k-means can now be called with options to determine preferred file format.

[0.1.2] - 2022-05-09

Fixed

• Fixed an issue with processing large datasets by moving from arrow files to arrows IPC format.

Changed

• Rebranded from clj-kmeans to josh.meanings, because I like the name meanings but figure I should namespace it.

[0.1.1] - 2022-05-06

Changed

• Dramatically improved performance on large datasets accomplished by switching from buffered csv reading to optimized tech.ml.dataset usage.

[0.1.0] - 2022-05-01

Added

• Initial project created with support for k-means clustering on larger than memory datasets.
• Only configuration provided is choice of k. Distance function used is earth mover distance.