Projection methods

A ProjectionConfig specifies an optional dimensionality reduction step applied to the transformed feature matrix before it is passed to a model. Setting projection_type to NONE skips this step entirely.


import mllabiome as mll
 
mll.ProjectionConfig(
    projection_type=mll.ProjectionType.PCA,
    n_components=50,
)

The n_components parameter determines the number of output dimensions. When set to None, the default for each method is used. The degree parameter applies only to POLYNOMIAL.

Available projection methods

Type	Description
`NONE`	No projection. The full transformed feature matrix is passed to the model.
`PCA`	Principal component analysis.
`INCREMENTAL_PCA`	Incremental PCA for datasets that do not fit in memory.
`KERNEL_PCA`	Kernel PCA, enabling non-linear projections via a kernel function.
`TRUNCATED_SVD`	Truncated singular value decomposition (LSA). Does not centre the data and is sparse-compatible.
`FAST_ICA`	Fast independent component analysis.
`SPARSE_PCA`	Sparse PCA, where components are constrained to be sparse.
`MINIBATCH_SPARSE_PCA`	Mini-batch variant of sparse PCA for larger datasets.
`NMF`	Non-negative matrix factorisation. Requires non-negative input.
`MINIBATCH_NMF`	Mini-batch variant of NMF.
`SPARSE_CODER`	Sparse coding against a pre-fitted dictionary.
`POLYNOMIAL`	Polynomial feature expansion. Use `degree` to control the polynomial degree.

Predefined configuration set

operations/projections_configs.py provides a ready-to-use list with one entry per projection type, each using its default number of components.


from operations.projections_configs import projections_configs
 
config = mll.ExperimentConfiguration(
    ...
    projection_configs=projections_configs,
    ...
)