Supervised Feature Selection via Ensemble Gradient Information from Sparse Neural Networks

Accepted by AISTATS 2024, by Kaiting Liu, Zahra Atashgahi, Ghada Sokar, Mykola Pechenizkiy, and Decebal Constantin Mocanu.

Abstract

GradEnFS is a novel resource-efficient supervised feature selection algorithm based on a sparse multi-layer perceptron. By utilizing gradient information from various sparse models across different training iterations, our method successfully identifies informative feature subsets.

Method

Usage

To initiate the program, please use the command "python main.py" along with the hyperparameters of your choice.

There are some main arguments:

--dataset(string): the dataset to be used.

--epsilon(int): hyperparameters for controlling the sparsity level.

--alpha(float): pruning rate during the topology update.

--beta(float): hyperparameter for the neuron importance metric.

To view all available hyperparameters and options, you can utilize the "python main.py --help" command.