Published version 1.0.0, including the following features:
- 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
andKLGEMINI
- 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
andgemclus.gemini.MMDOvO
are nowgemclus.gemini.MMDGEMINI
gemclus.gemini.WassersteinOvA
andgemclus.gemini.WassersteinOvO
are nowgemclus.WassersteinGEMINI
- Both the MMD and Wasserstein GEMINI now support custom kernel/metric parameters
- Fixing a gradient mistake in the
gemclus.MI