Methodology and interpretationComparison design
- Local HUGIML results are joined to published TabArena results by exact dataset name and outer-split number.
- Binary error is 1 − ROC AUC; multiclass error is log loss. Split errors are averaged within dataset before aggregation, giving every dataset equal weight.
- Each tab is an independent comparison pool. Elo, ranks, normalized score, improvability, confidence intervals, and HUGIML pairwise results are recomputed within that pool.
- AutoMLPipelineFeatureGenerator is fitted separately on every inner child training fold and retained with that child for validation and outer-test transformation; no validation or test rows contribute to preprocessing state.
- The outer tests and metrics align. HUGIML evaluates 32 configurations, while each tuned official baseline evaluates 200; retained ensembles and compute budgets remain method-specific.
- Detailed metric tables show mean, median, and the paired HUGIML advantage across accuracy, balanced accuracy, F1, ROC AUC, and log loss. Positive deltas favor HUGIML; log loss uses the lower-is-better direction.
HUGIML hyperparameters
32 configurations per outer split; eight retained inner-fold models per selected configuration. Binary selection uses ROC AUC; multiclass selection uses log loss. Selected parameters are recorded for each dataset and outer split.
| Parameter | Search values |
|---|
| B | -1 |
| adaptive_binning | true |
| L | 1, 2 |
| topK | 50, 100, 150, 200 |
| feature_mode | original_plus_patterns |
| G | 0.01, 0.001 |
| convert_binary_to_categorical | false |
| augmented_pair_transforms | true |
| topk_budget_strict | false |
| lr_solver | adaptive_l1 |
| lr_source_policy | main_effect |
| lr_C | 0.5 |
| base_estimator | LR (automatic sparse logistic regression); One-vs-Rest LeafWiseBoundedLookaheadRPTEFeatureLR (lr_C=0.5, lr_penalty=l1) |