kmodes

What's changed

• Improve estimation of gamma for k-prototypes (https://github.com/nicodv/kmodes/pull/186)

Full Changelog: https://github.com/nicodv/kmodes/compare/0.12.1...0.12.2

What's changed

• Fix for broken fit_predict on KPrototypes (https://github.com/nicodv/kmodes/pull/176)
• Improved validation of sample weights (https://github.com/nicodv/kmodes/pull/176)

Full Changelog: https://github.com/nicodv/kmodes/compare/0.12.0...0.12.1

What's changed

• Support for sample weights for both k-modes and k-prototypes algorithms, courtesy of @kklein (https://github.com/nicodv/kmodes/pull/174, https://github.com/nicodv/kmodes/pull/171)
• Add official support for Python 3.10 (https://github.com/nicodv/kmodes/pull/170)
• Bugfix for algorithm convergence (https://github.com/nicodv/kmodes/commit/370d64b1067331b413d641103a52bd4c636ac702)
• Switch internally to pytest from nose (https://github.com/nicodv/kmodes/pull/170)
• Some small fixes and cleanups

Full Changelog: https://github.com/nicodv/kmodes/compare/0.11.1...0.12.0

What's Changed

• 155: Make _labels_cost function public by @nicodv in https://github.com/nicodv/kmodes/pull/156
• Iterations were running for 1 more than expected by @nicodv in https://github.com/nicodv/kmodes/pull/160
• Change feature array initialization dtype to uint32 by @rggelles in https://github.com/nicodv/kmodes/pull/166. This reduces memory footprint significantly.
• Drop support for missing values, following scikit-learn: https://github.com/nicodv/kmodes/commit/a20f6ed6567f4c0d5c5c9ad70ca86a6b77ab522f

Full Changelog: https://github.com/nicodv/kmodes/compare/0.11.0...0.11.1

• Python 3.9 support
• Minimum sklearn version upgrade to 0.22
• Default init method for k-prototypes is now the Cao method (same as k-modes and in line with documentation), courtesy of @larroy
• Optimizations

• Added Jaccard dissimilarity function, courtesy of @BikashPandey17 (#129 )
• Return the costs per epoch after training, courtesy of @daffidwilde (#79 )
• Python 3.8 now supported
• Python 3.4 no longer supported because sklearn dropped it too
• Various bugfixes and improvements

• Bugfixes

• Categorical variables are now automatically encoded and decoded between original data values and integers (used internally by k-modes). User does not have to use to the categorical variable mapping anymore when looking at the cluster centroids.
• Support for custom dissimilarity measures
• Python 3.6 support
• More robust manual initialization

• Huge speedup for k-prototypes, especially for large numbers of samples (#45). A k-prototypes benchmark script is included in examples now.
• Offer an implementation of Ng's dissimilarity measure, which could improve convergence (#37).
• Allow pandas DataFrames to be presented to the algorithm, instead of just numpy arrays (#40).
• Improved handling of dependencies (#49, #53).
• Various small bugfixes and improvements.

• Support for more than 256 clusters
• Optional parallel execution of the multiple initialization runs (courtesy of @rphes )
• Enhanced error checking when using pandas DataFrames as inputs to the algorithms
• Various bug fixes and improvements
• Semantic versioning from now on

• Improved pandas compatibility, courtesy of @Genie-Liu
• Forward compatible with future scikit-learn versions that will no longer include joblib, courtesy of @trevorstephens