Feature ranking for semi-supervised learning

Abstract The data used for analysis are becoming increasingly complex along several directions: high dimensionality, number of examples and availability labels the examples. This poses a variety challenges existing machine learning methods, related to analyzing datasets with large that described in high-dimensional space, where not all have provided. For example, when investigating toxicity chemical compounds, there many compounds available can be information-rich representations, but information on their toxicity. To address these challenges, we propose methods semi-supervised (SSL) feature rankings. rankings learned context classification regression, as well structured output prediction (multi-label classification, MLC, hierarchical multi-label HMLC multi-target MTR) tasks. is first work treats task ranking uniformly across various tasks prediction. best our knowledge, it also SSL MTR. More specifically, two approaches—based predictive clustering tree ensembles Relief family algorithms—and evaluate performance 38 benchmark datasets. extensive evaluation reveals based Random Forest perform (incl. MLC tasks) fastest tasks, while extremely randomized trees regression Semi-supervised outperform supervised counterparts majority different showing benefit using unlabeled addition labeled data.