Abstract
In unsupervised settings, where labeled data is unavailable, identifying key features becomes more difficult but remains crucial. While various unsupervised feature selection methods have been developed recently, there is still considerable room for improvement. In this paper, we present a novel unsupervised feature selection algorithm that introduces Variable Priority (VarPro) in an unsupervised context, which we call Unsupervised Variable Priority (UVarPro). We adapt VarPro from the supervised to the unsupervised domain by reframing the problem as a localized two-class classification analysis, where class labels correspond to regions defined by specific rules. The approach is supported by Bayes’ theorem. Local regression analysis is incorporated as an extra layer of supervision at the rule level, helping to reduce dimensionality and mitigate the impact of noisy variables in high-dimensional data. In extensive experiments on synthetic and real-world datasets across classification, regression, and survival tasks—evaluated with multiple metrics and supervised gold standards—UVarPro consistently selects more informative features, outperforms state-of-the-art unsupervised methods in clustering accuracy, and enhances visualization performance.