Random Feature Selection Outperforms Many Unsupervised Methods
A new arXiv preprint (2605.22973) argues that many state-of-the-art unsupervised feature selection methods fail to outperform random feature selection. The authors propose using random selection as a baseline for evaluation, showing empirically that numerous published methods are less effective and less efficient than a random baseline. They stress the need for this baseline in future development to ensure consistent progress.
Key facts
- arXiv:2605.22973 proposes random feature selection as a baseline for unsupervised feature selection.
- Many state-of-the-art methods are outperformed by random selection in both performance and efficiency.
- The authors emphasize the strict requirement of considering random feature selection as a baseline.
- The study is based on empirical evaluation on selected datasets.
- The paper is categorized as a cross-type announcement on arXiv.
Entities
Institutions
- arXiv