Additional UMAP interfaces

Squeeze retains UMAP-family functionality inherited from umap-learn. These interfaces are distinct from the new Rust algorithms and are not part of the 11-method heatmap comparison.

Density-aware UMAP

UMAP(densmap=True) adds density-oriented optimization. Relevant controls are dens_lambda=2.0, dens_frac=0.3, and dens_var_shift=0.1. output_dens=True changes the fit-transform result to (embedding, radii_original, radii_embedding); the radii are log-transformed. It can also be enabled without densMAP. A tuple return is therefore possible instead of the ordinary coordinate array. densMAP does not support the standard UMAP out-of-sample transform path.

AlignedUMAP

AlignedUMAP fits a list of related datasets. Its fit and fit_transform calls require relations=[mapping_0_to_1, mapping_1_to_2, ...], with each mapping a dictionary of corresponding row indices in consecutive datasets. There must be one fewer relation than datasets. alignment_regularisation controls coupling; alignment_window_size controls the alignment window. The result is a list of embeddings, not a single matrix. update supports adding a related dataset; there is no ordinary independent-row transform method.

ParametricUMAP

ParametricUMAP uses TensorFlow/Keras to train a neural encoder. The optional parametric dependency is required; without it, construction raises ImportError. The module also imports additional optional training/export packages, so availability must be checked in the target environment. This path has not been validated by the core documentation smoke run and is not a Rust or PyTorch backend.

Its constructor takes batch_size, dims, encoder, decoder, parametric_reconstruction, reconstruction-loss options, global-correlation and landmark-loss options, keras_fit_kwargs, and inherited UMAP keyword arguments. It provides model saving through .save(...) and loading via squeeze.parametric_umap.load_ParametricUMAP(...).

Precomputed neighbors

UMAP(precomputed_knn=(indices, distances, search_index)) accepts a compatible neighbor graph. Indices and distances must match input row order and neighborhood size. New-row transformation requires an appropriate search index, not merely the two arrays. Retain graph construction parameters and provenance alongside the data.

See the implementation docstrings for detailed mode restrictions. These optional interfaces should be tested against the intended workload before deployment.