Replication data for: Fitness and Cost Landscapes for Evolutionary Computation: Experiments in Feature Selection for Classification

This README file was generated on [2026-04-18] (YYYY-MM-DD). Last updated: [2026-04-18]. ------------------- GENERAL INFORMATION ------------------- // Title of Dataset: Replication data for: Fitness and Cost Landscapes for Evolutionary Computation: Experiments in Feature Selection for Classification // DOI: 10.5281/zenodo.19636954 // Data Type: Numeric data as .h5 files. // Date of data collection/generation: 2026-03-05 (YYYY-MM-DD). // Description of dataset: This dataset contains replication data for the paper Fitness and Cost Landscapes for Evolutionary Computation: Experiments in Feature Selection for Classification. It contains accuracy and training cost tables of well-known classification datasets from The UCI Machine Learning Repository. These tables are intended to be used as surrogates for fitness and cost landscapes of feature selection optimization problems, and they comprise the accuracy for all feature subsets, i.e., all column combinations, obtained by five classifiers: 1. k-Nearest Neighbors (kNN) 2. Logistic Regression (LR) 3. Neural Network (NN) 4. Random Forest (RF) 5. Support Vector Machine (SVM) -------------------------- METHODOLOGICAL INFORMATION -------------------------- // Description of sources and methods used for collection/generation of data: We study 10 datasets from the University of California, Irvine (UCI) ML Repository. On each dataset, we train all classifiers and compute an accuracy table that comprises all possible 2^n subsets of features, where n is the number of columns in the dataset. We provide 32 repetitions per feature subset, to address the stochasticity of the ML models. // Methods for processing the data: The tables were computed using Julia 1.12.2, using the following packages: 1. k-Nearest Neighbors (KNNClassifier from NearestNeighbors.jl and its MLJ.jl interface) 2. Logistic Regression (MultinomialClassifier from MLJLinearModels.jl) 3. Neural Network (MLP from Flux.jl and its MLJ.jl interface) 4. Random Forest (RandomForestClassifier from BetaML.jl and its MLJ.jl interface) 5. Support Vector Machine (SVC from LIBSVM.jl and its MLJ.jl interface) In all models, the default parameters (as of MLJ v0.21.0) were used, with a 70/30 train/test split. // Facility-, instrument- or software-specific information needed to interpret the data: The data can be read using any data analysis tool that can read HDF5 (.h5) files. -------------------- DATA & FILE OVERVIEW -------------------- // File List: Each of 10 ML datasets x 5 ML classifiers is stored as a separate .h5 file, with a dataset ID as the filename. Each file contains two HDF5 datasets: 1. accuracies: the accuracy of a given classifier trained on a given feature subset of a given dataset 2. times: the time it took to train a given classifier trained on a given feature subset of a given dataset Each HDF5 dataset is a matrix of size 2^n x 32, with ordered rows (from 1 to 2^n-1 possible feature subsets). Columns represent each of 32 repetitions. The i-th row in a matrix is the decimal representation of each of the bitstrings representing a feature subset, where each digit in the bitstring corresponds to a column in the ML dataset, with a 1 or 0 depending whether or not it is used for training the model. The cell contains a real number for the ML classifier accuracy using that feature subset, or the time (in seconds) used to compute the model. Available datasets: i Dataset Name File Name n m Source 1 breast-cancer-wisconsin-original 01-breast-w 9 699 https://archive.ics.uci.edu/dataset/15/breast+cancer+wisconsin+original 2 glass-identification 02-glass 9 214 https://archive.ics.uci.edu/dataset/42/glass+identification 3 heart-disease-cleveland 03-heart-c 13 303 https://archive.ics.uci.edu/dataset/45/heart+disease 4 wine 04-wine 13 178 https://archive.ics.uci.edu/dataset/109/wine 5 credit-approval 05-credit-a 15 690 https://archive.ics.uci.edu/dataset/27/credit+approval 6 zoo 06-zoo 16 101 https://archive.ics.uci.edu/dataset/111/zoo 7 congressional-voting-records 07-congress-v 16 435 https://archive.ics.uci.edu/dataset/105/congressional+voting+records 8 letter-recognition 08-letter-r 16 20000 https://archive.ics.uci.edu/dataset/59/letter+recognition 9 seismic-bumps 09-seismic-b 18 2584 https://archive.ics.uci.edu/dataset/266/seismic+bumps 10 hepatitis 10-hepatitis 19 155 https://archive.ics.uci.edu/dataset/46/hepatitis Where n is the number of features and m is the number of training examples. The original dataset can be consulted at the UCI Machine Learning Repository, provided in the Source column.