Phase Transitions in Unsupervised Feature Selection

Dataset description This dataset contains all the data required to reproduce the analyses presented in: Fiorentino J., Monti M., Vrachnos D. M., Del Tatto V., Laio A., Tartaglia G. G., Phase Transitions in Unsupervised Feature Selection. arXiv, 2026 https://doi.org/10.48550/arXiv.2602.00660 The data support a systematic study of backward feature elimination using the Differentiable Information Imbalance (DII) across multiple protein datasets, feature types, and controlled correlation regimes. Data organization The DATA folder is organized as follows: DATASETS/ Physico-chemical (82 features) and structural (67 features) descriptors computed for the human proteome Train and test datasets for four protein classes: LLPS (liquid–liquid phase separating proteins) RBP (RNA-binding proteins) MEMBRANE proteins ENZYME proteins Separate datasets are provided for physico-chemical and structural feature sets Subsamples/ Random balanced subsamples of increasing size Used to study scaling properties and phase-transition-like behavior in backward feature elimination Each subsample is generated independently and consistently across feature sets and protein classes DIIvsF/ Values of the Differentiable Information Imbalance (DII) during the backward feature elimination process Reported for both training and test subsamples Provided for all four protein datasets weightsvsF/ Optimal feature weights identified by the DII at each step of the backward feature elimination process Reported for all four protein datasets and feature sets LassoFull/ Results of backward feature elimination using the DII applied to the full protein datasets Enables comparison between subsampled and full-dataset regimes CorrelationTuning/ Results of backward feature elimination using the DII under controlled feature statistics: Feature correlation tuned via parameter β ∈ [−1, 1] Feature variance heterogeneity tuned via parameter α ∈ [0, 1] at β = −1 (complete feature decorrelation) Used to disentangle the role of correlations and variance heterogeneity in the observed phase transition BinaryClassification/ Binary classification performance metrics obtained by training a multilayer perceptron (MLP) Classifiers are trained using features selected by the DII at each step of the backward feature elimination process Results are provided for: all random subsamples all four protein datasets all feature elimination stages Reproducibility All data are designed to be used in conjunction with the accompanying GitHub repository, which provides: the full analysis code, environment specifications, and a Jupyter notebook to reproduce all figures and results. License This dataset is released under the Creative Commons Attribution 4.0 (CC BY 4.0) license.