Supplementary material for the paper: Objective-Induced Bias and Search Dynamics in Multiobjective Unsupervised Feature Selection

This repository contains supplementary material for our PPSN submission, including the Python code used to generate the synthetic benchmark dataset and supplementary Figures from the following multiobejctive formulations: maximise silhouette - maximise subset-size maximise silhouette - minimise subset-size maximise accuracy - maximise subset-size maximise accuracy - minimise subset-size minimise PCA loss - maximise subset-size minimise PCA loss - minimise subset-size Each formulation has 5 different Figures: search: the search history, initial population and achieved front hvhist: the hypervolume convergence composition: the clustermap of selected features across the front frontclust: the visualisation of the found clusters on the front (colored) boxclust: the test accuracy distribution of the solutions in each cluster Each of these Figures have three panels: one for each investigated initialisation strategy. File names are formated in the following way: {type}_{objective 1}_{objective 2}. The code used for generating the synthetic dataset can be found in generate_data.py The presented Figures have been obtained with the same optimisation parameters and dataset presented in the paper.