Evaluating machine leaning algorithms for accuracy, stability, and among-predictors discriminability in modeling species-richness across ten datasets

Global biodiversity is experiencing substantial declines, and mitigating this crisis requires analytical approaches that can accurately predict biodiversity in relation to natural conditions and human-induced stressors. While numerous machine learning (ML) algorithms for regression are available for such analyses, synthesizing outcomes across studies is challenging due to: (1) reliance on single datasets, limiting generalizability; (2) varying modeling processes; (3) inconsistent performance criteria; and (4) limited consideration of model stability and among-predictor discriminability. We addressed these issues by applying five ML algorithms—Random Forest (RF), Boosted Regression Tree (BRT), Extreme Gradient Boosting (XGB), Conditional Inference Forest (CIF), and Lasso—to ten large datasets on freshwater fish, mussels, and caddisflies. Using consistent modeling methods, we evaluated accuracy (R 2 and RMSE), stability (coefficient of variation of R 2 and RMSE), and among-predictors discriminability (variation in predictor importance). RF, BRT, and XGB generally achieved higher accuracy than CIF and Lasso, although performance varied by dataset. CIF, however, was the most stable (average CoV-R 2 = 0.12), followed by RF, XGB, and BRT (0.13–0.15). BRT was most effective at distinguishing among predictors, followed by CIF and Lasso. Considering all criteria, CIF, XGB, and BRT ranked similarly high, followed by RF and Lasso. The top three models also showed similar predictor rankings, while RF and Lasso differed. Reducing predictors by 58 % had little effect on accuracy or stability, and averaging predictions across replicate models should mitigate the effects of model stability. These findings support more robust ML applications in biodiversity research. • Tree-based models (XGB, BRT, RF, CIF) outperformed Lasso regression in accuracy. • All models showed low to moderate stability, although CIF was most stable. • BRT discriminated predictors for importance best, followed by CIF and Lasso. • XGB, BRT, and CIF ranked predictors similarly; RF and Lasso differed. • Predictor selection had minor impacts on the model performances.