OpenML Benchmark
Tuned DiRe achieved topology error 0 on all 11 OpenML datasets. At least one Pareto-optimal DiRe configuration dominated cuML UMAP on both kNN accuracy and topology error in 7 of 11 datasets. In the covertype case, all three Pareto-optimal…
1 sources - 5 claims
Tuned DiRe achieved topology error 0 on all 11 OpenML datasets. At least one Pareto-optimal DiRe configuration dominated cuML UMAP on both kNN accuracy and topology error in 7 of 11 datasets. In the covertype case, all three Pareto-optimal DiRe trials strictly dominated cuML UMAP. The OpenML benchmark used NSGA-II through Optuna with objectives to maximize kNN accuracy and minimize topology error. The Pareto-optimal hyperparameters converged on spectral initialization, larger spread, moderate-to-high iterations, and 10 to 30 neighbors.