Hugging Face's logo

Datasets:
TabQueryBench2026
/
server

Dataset card Files
xet
 Community
1
server / evaluation /query_family /subgroup /support3_analysis /README.md
TabQueryBench's picture
TabQueryBench
Upload subgroup support3 analysis
e6c15ac verified
preview code
|
Raw
History Blame Contribute Delete
3.27 kB

Support3 analysis

This directory reframes the subgroup comparison without using the original weighted composite query score from real_panel_experiment.py.

Why support3_score

We define:

support3_score = mean(key_set_score, row_count_score, column_score)

This is the cleanest lightweight proxy for subgroup support preservation because it directly asks:

  • did the same subgroup keys appear?
  • did a similar number of grouped rows survive?
  • did the output keep the same column scaffold?

It avoids:

  • the original hand-set weights
  • strict_set_score, which is dominated by exact value matching
  • the ambiguous use of profile_score in the implementation, which is really a key-column marginal score rather than a subgroup analytical-value score

What survives after removing the original weights

The core claim still survives:

subgroup_size_stability is better preserved than internal_profile_stability in a support-preservation sense.

At the query level:

  • internal mean: 0.755
  • size mean: 0.818
  • mean gap: +0.063 in favor of size

At the panel level:

  • internal mean: 0.754
  • size mean: 0.810
  • mean gap: +0.055 in favor of size

Can we also say size changes less?

Not in every possible sense.

Across the three regimes (c, m, n), the absolute mean range is not smaller for size:

  • internal panel means: 0.759 -> 0.818 -> 0.709, range 0.110
  • size panel means: 0.811 -> 0.881 -> 0.752, range 0.129

So the strict statement:

"size has smaller regime-to-regime movement"

is not the safest one.

The safer and better-supported statement is:

size has a lighter low-score tail and lower relative dispersion under support3_score.

Evidence:

  • panel-level CV:
    • internal: 0.265
    • size: 0.246
  • panel-level median absolute deviation from the median:
    • internal: 0.180
    • size: 0.102
  • panel share with support3 <= 0.4:
    • internal: 5.8%
    • size: 2.7%
  • panel share with support3 <= 0.8:
    • internal: 54.2%
    • size: 38.3%

So the stronger reading is:

size is not necessarily flatter across all regimes, but it is more resistant to severe degradation.

Files