Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]
Supervised Feature Selection with Neuron Evolution in Sparse Neural Networks
Authors: Zahra Atashgahi, Xuhao Zhang, Neil Kichler, Shiwei Liu, Lu Yin, Mykola Pechenizkiy, Raymond Veldhuis, Decebal Constantin Mocanu
TMLR 2023 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | By performing several experiments on 11 low and high-dimensional real-world benchmarks of different types, we demonstrate that Neuro FS achieves the highest ranking-based score among the considered state-of-the-art supervised feature selection models. |
| Researcher Affiliation | Academia | 1 Department of Electrical Engineering, Mathematics and Computer Science, University of Twente, The Netherlands 2 Department of Mathematics and Computer Science, Eindhoven University of Technology, The Netherlands |
| Pseudocode | Yes | Neuro FS is schematically described in Figure 1 and the corresponding pseudocode is available at Algorithm 1. |
| Open Source Code | Yes | The code is available on Git Hub1. 1https://github.com/zahraatashgahi/Neuro FS |
| Open Datasets | Yes | Datasets. We evaluate the effectiveness of Neuro FS on eleven datasets2 described in Table 1. 2Available at https://jundongl.github.io/scikit-feature/datasets.html |
| Dataset Splits | Yes | Table 1: Datasets characteristics. ... # Train # Test ... Then, we train a SVM classifier on the selected features subset of the training set. We report the classification accuracy on the test set as a measure of performance. |
| Hardware Specification | Yes | We used a Dell R730 processor to run the experiments. We run neural network-based methods using Tesla-P100 GPU with 16G memory. |
| Software Dependencies | No | We implemented our proposed method using Keras Chollet et al. (2015). ... For Fisher_score, CIFE, ICAP, and RFS, we have used the implementations provided by the Scikit-Feature library Li et al. (2018)4. ... we use classification accuracy of a SVM classifier Keerthi et al. (2001) with RBF kernel implemented by Scikit-Learn library7 (The paper mentions software like Keras and Scikit-Feature library but does not provide specific version numbers for reproducibility.) |
| Experiment Setup | Yes | Hyperparameters. The architecture of the network used in the experiments is a 3-layer sparse MLP with 1000 neurons in each hidden layer. The activation function used for the hidden layers is Tanh (except for Isolet dataset where Relu is used), and the output layer activation function is Softmax. The values for the hyperparameters were found through a grid search among a small set of values. We have used stochastic gradient descent (SGD) with a momentum of 0.9 as the optimizer. The parameters for training neural network-based methods, including batch size, learning rate, and the number of epochs (tmax), have been set to 100, 0.01, and 100, respectively. However, the batch size for datasets with few samples (m <= 200) was set to 20. The hyperparameter determining the sparsity level ε is set to 30. Update fraction for the input layer ζin and hidden layer ζh have been set to 0.2 and 0.3 respectively. Neuron removal duration hyperparameter α is set to 0.65. ζin and α are the only hyperparameters particular to Neuro FS. We use min-max scaling for data preprocessing for all methods except for the BASEHOCK dataset, where we perform standard scaling with zero mean and unit variance. |