Ensemble Diversity
Overview
Ensemble diversity concerns whether combining models that make different errors improves predictions over relying on the strongest model alone. Researchers measure disagreement and error correlation, then use voting, weighted selection, or stacking to combine members; other approaches encourage diversity through architectural variation, shared-weight ensembles, residual methods, or deliberation. The work spans image classification and out-of-distribution detection, tabular prediction, subjective text labeling, forecasting, and multi-agent language-model tasks. A central tension is that diversity is useful only when members are capable and aggregation is effective, while common diversity metrics can confound complementarity with model skill and gains in accuracy may come at the cost of calibration.
Latest papers
23 papers3 papers in the last four weeks, level with the four weeks before. 0.0% of all new papers.