From Numerical Taxonomy to Classifier Modeling: A Quantitative Taxonomic Workflow for Euterpnosia Cicadas (Hemiptera:...

Numerical taxonomy reveals morphological structure, whereas classifier modeling tests identification against labeled reference hypotheses. We evaluated 70 Taiwanese Euterpnosia Matsumura, 1917 specimens representing five operational classes, using 71 external characters for unsupervised analysis and 20 non-destructive characters for supervised modeling. Broad character retention was separated from goal-specific selection: taxonomists may prespecify candidates from literature or experience, while data-driven selection remained inside training folds. In 20 × five-fold nested cross-validation, feature-screened multinomial accuracy was 95.93% (balanced accuracy 94.71%); the all-character model reached 96.57%, showing that selection need not force parsimony when a compact pool is already informative. Leave-one-species-out tests rejected omitted E. chilanensis Chen, Hsieh, Chen, Chen & Chang, 2021, E. olivacea Kato, 1927, and E. hoppo Matsumura, 1917 in 100%, 100%, and 92.0% of decisions, but E. alpina Chen, 2005 and E. varicolor Kato, 1926 only 21.1% and 35.0%. CART selected X49, X20, and X32 for a concise quantitative-key draft. The workflow can prioritize candidate diagnostic characters and support identification, abstention, and key construction, but does not independently establish species boundaries or nomenclatural conclusions.