Adaptive Unsupervised Feature Selection with Hierarchical Granularity Optimization

With the rapid advancement of AI and sensor technologies, high-dimensional unlabeled data are increasingly common in areas like healthcare and the Internet of Things. Unsupervised feature selection is crucial for reducing dimensionality while preserving information, yet existing multi-scale methods often use fixed parameters, limiting their adaptability to hierarchical data structures and dynamic feature relationships. To overcome this, we propose a Hierarchical Granularity-Optimized Feature Selection (HGUMFS) algorithm. It uses the Hopkins statistic to dynamically determine data structure and constructs a recursive spectral clustering hierarchy, along with a dual-modal fuzzy similarity mechanism and decay-weighted feature fusion to identify optimal subsets. Evaluations on public datasets show that HGUMFS improves clustering accuracy and feature reduction rates over existing methods.