evaluation/query_family/subgroup/analysis_report.md

Subgroup Breakdown Report

Scope

• Source analysis run: 20260526_v2_official20_plus_c2patch_plus_subgroupconditionalrepair_merged49
• Family analyzed: subgroup_structure
• Excluded models: cdtd, codi, goggle
• Included models: 12
• Deduplicated dataset-model panels: 559
• Subgroup query rows used: 12515

Canonical decomposition

• subgroup_structure = 0.5 * internal_profile_stability + 0.5 * subgroup_size_stability
• internal_profile_stability captures subgroup-internal feature/distribution behavior.
• subgroup_size_stability captures whether subgroup support/size structure is preserved.

Main findings

1. REAL is the expected perfect upper bound with subgroup score 1.000. Among synthetic generators, RealTabFormer is strongest with mean subgroup score 0.790 across 49 datasets.
2. BayesNet leans most toward internal-profile preservation (profile minus size = 0.002), while ForestDiffusion is the clearest size-heavy model (profile minus size = -0.033).
3. TabDDPM is the most balanced model between the two subgroup branches with mean absolute branch gap 0.000.
4. Dataset difficulty is uneven: n9 is the hardest dataset on subgroup score (0.091 mean across models), while c3 is the easiest (1.000).

Files to use first

• figures/subgroup_tradeoff_scatter_main.pdf
• figures/subgroup_branch_dumbbell_main.pdf
• figures/subgroup_prefix_bars_appendix.pdf
• tables/subgroup_model_summary_generated.tex
• data/model_summary.csv

Prefix note

• Prefix coverage summary rows: 36
• The c / m / n split is exported explicitly because subgroup behavior differs by dataset family, not just by overall model average.