server / missingness_breakdown /v2 /analysis_report__v2.md
TabQueryBench's picture
Add files using upload-large-folder tool
8ee4e88 verified
|
Raw
History Blame
1.04 kB
# Missingness Breakdown
- Source mode: `direct_missingness_evaluator`
- Reference analysis run: `trainonly_v2_current_success_official_20way_official20_20260519_232817`
- Included models: `11`
- Deduplicated dataset-model panels: `509`
- Direct model-dataset rows used: `509`
- Direct target rows used: `3825`
## Canonical decomposition
- `missingness_structure = 0.5 * marginal_missing_rate_consistency + 0.5 * co_missingness_pattern_consistency`
- Canonical `co_missingness_pattern_consistency` now uses profile-only edge averaging.
- `co_missing_strength_score` is exported separately as an auxiliary diagnostic and is not folded into the two-subitem family score.
- `co_missing_composite_score` preserves the previous 0.7-profile / 0.3-strength blend for sensitivity analysis only.
- This bundle bypasses SQL query analysis and uses the canonical direct co-missing evaluator over real-train vs synthetic CSV pairs.
- The standardized appendix bundle now mirrors tradeoff, prefix, and heatmap views just like the other query families.