Unsupervized Feature Selection

Unofficial implementation of the unsupervised feature selection algorithm proposed by Ono in March 2020 [1].

Installation

1. Just add the fs_ono2020.py to your directory.
2. Import fs_ono2020.FeatureSelector class.

Usage of FeatureSelector

See demo.ipynb.

parameters

• n_features

int
Number of features to be selected.

• random_state

int (optional, default: None)
Specify integer if you want reproducible output.

• logging

bool (optional, default: False)
Specify True if you want to see the progress of instance fitting.

• loop_limit

positive int (optional, default: numpy.inf)
Specify positive integer if you want the optimization loop to end in the middle.

methods

• fit(X)

Fit the FeatureSelector instance with your numpy.ndarray dataset X.

• fit_transform(X)

Fit the FeatureSelector instance with your numpy.ndarray dataset X, and return new dataset X_selected with the selected features.

• transform(X)

Return new dataset X_selected with the selected features.

• reconstruct(X_selected)

Return the reconstructed dataset X_reconstructed from the transformed dataset X_selected.

attributes

• selected

The list of selected feature indices

• deselected

The list of deselected feature indices

• original_dim

The original dimension of the input vectors

Algorithm

In the abstruct of the paper:

"In this study, we consider an objective function defined as the reconstruction loss of a linear autoencoder, and this is formulated as a discrete optimization problem that selects the element that minimizes it. Also, we propose a method to solve this problem by sequentially replacing elements chosen so that the objective function becomes smaller."

Reference

[1] Nobutaka Ono, "機械学習における乗算を用いない次元削減(Dimension reduction without multiplication in machine learning)," IEICE Tech. Rep., vol.119, no.440, SIP2019-106, pp.21-26, March 2020. (in Japanese). URL: https://www.ieice.org/ken/paper/20200302U1Xv/eng/