RELEASES¶

History of changes¶

Adding the kernelised version of RIM with: KernelRIM
Adding the dynamic version of paths for feature selection in sparse models. A simply argument dynamic=True activates the dynamic mode.
Possibility of passing custom kernels and metrics to sparse models. This is not compatible with the dynamic mode.
No need to specify any longer the full partition of the features in the groups arguments of the sparse models
New GEMINIs: HellingerGEMINI, TVGEMINI and KLGEMINI
Introducing generic models that can be combined with any GEMINI
- gemclus.linear.LinearModel, gemclus.mlp.MLPModel, gemclus.nonparametric.CategoricalModel, gemclus.sparse.SparseLinearModel, gemclus.sparse.SparseMLPModel
- The GEMINI parametrisation of DOUGLAS can now be done through string
- The dedicated MMD and Wasserstein models remain and support custom kernel/metric parameters
Fusing GEMINIs into a single class per distance
- gemclus.gemini.MMDOvA and gemclus.gemini.MMDOvO are now gemclus.gemini.MMDGEMINI
- gemclus.gemini.WassersteinOvA and gemclus.gemini.WassersteinOvO are now gemclus.WassersteinGEMINI
- Both the MMD and Wasserstein GEMINI now support custom kernel/metric parameters
Fixing a gradient mistake in the gemclus.MI

Adding a new sparse logistic regression model trained with mutual information instead of MMD GEMINI: gemclus.sparse.SparseLinearMI
Adding new package containing methods for unsupervised tree clustering with end-to-end training: gemclus.tree. The package features a CART-like algorithm for clustering named Kauri and a differentiable tree named Douglas
Experimental: A method for adding constraints of type must-link cannot-link to discriminative models: gemclus.add_mlcl_constraint
Minor fixes in documentation
Better compatibility with scikit learn 1.3.0 regarding parameter constraint check

Fixing the ABCMeta parameter validation problem for the draw_gmm method for retrocompatibility with Python 3.8.
Constraining the package to Python>=3.8 to respect the requirements of the package.
Minor fix on the get_selection method for the Linear sparse models to respect the 1d output shape of the array.

Isolating the definition of GEMINIs in a separate classes for external usages: gemini.MMDOvO, gemini.WassersteinOvA etc.
Adding the nonparametric models in package gemclus.nonparametric with 2 additional examples for its usage in graph node clustering.
Fixing control variables in the path method for spars models.

Adding the Gaussian+Student-t mixture dataset: gstm
Method for sampling multivariate student-t distributions: multivariate_student_t
Adding tests for data package
Adding the possibility of a precomputed kernel/distance passed to fit
Adding batch size parameters
Fixing zero division in sparse linear model proximal gradient