kerneloose 0.0.2

Kernel method for out-of-sample extension

kerneloose

Python implementation of the kernel method for out-of-sample extension (OOSE) of dimensionality reduction techniques.

Based on "Parametric nonlinear dimensionality reduction using kernel t-SNE" by Gisbrecht, Schulz, and Hammer.

The kernel method is particularly useful for projection techniques that are computationally expensive and/or have non-convex objective functions, such as t-SNE.

Installation

pip install kerneloose

Usage example

The syntax follows scikit learn conventions. Assume hd_data is a numpy array containing high-dimensional data, and an array ld_data of equal length but lower dimension was obtained by some projection technique. An OOSE of that projection can be obtained by:

from kerneloose import KernelMap

kernel_oose = KernelMap()
kernel_oose.fit(hd_data, ld_data)

The mapping can be applied to new_data (with same dimensionality as ld_data) simply by:

kernel_oose.transform(new_data)

Parameters of the calculated OOSE mapping can be saved and loaded for later use:

kernel_oose.save('some/file/name')

resume_later = KernelMap()
resume_later.load('some/file/name')