File size: 13,892 Bytes
42dd317 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 | from __future__ import annotations
from pathlib import Path
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
ROOT = Path(__file__).resolve().parents[4]
BASE = ROOT / "Evaluation" / "query_fivepart_breakdown" / "conditional_breakdown"
OUT_DIR = BASE / "strength_focus"
RUN_DIR = (
BASE
/ "locality_support_diagnostics"
/ "runs"
/ "20260502_064421_conditional_locality_support"
/ "data"
)
MODEL_COLORS = {
"RealTabFormer": "#332288",
"TVAE": "#4477AA",
"ForestDiffusion": "#228833",
"TabDDPM": "#EE7733",
"TabSyn": "#66CCEE",
"TabDiff": "#AA3377",
"CTGAN": "#EE6677",
"ARF": "#777777",
"BayesNet": "#CCBB44",
"TabPFGen": "#009988",
"TabbyFlow": "#882255",
}
SUPPORT_BUCKET_ORDER = ["dense", "medium", "sparse"]
SUPPORT_BUCKET_LABELS = {"dense": "Dense", "medium": "Medium", "sparse": "Sparse"}
SUPPORT_BUCKET_COLORS = {"dense": "#1b9e77", "medium": "#7570b3", "sparse": "#d95f02"}
def _assign_primary_support_buckets(audit_df: pd.DataFrame) -> pd.DataFrame:
case_df = audit_df[
[
"dataset_id",
"query_id",
"real_support_value",
"support_main_eligible",
"support_recovery_mode",
"template_name",
]
].drop_duplicates()
eligible = case_df[
(case_df["support_main_eligible"] == True)
& (case_df["support_recovery_mode"].isin(["exact", "derived_exact"]))
& (case_df["real_support_value"].notna())
].copy()
rows: list[pd.DataFrame] = []
for dataset_id, group in eligible.groupby("dataset_id", sort=False):
values = pd.to_numeric(group["real_support_value"], errors="coerce")
if group.shape[0] < 3 or values.dropna().nunique() < 3:
continue
ranked = values.rank(method="first")
bins = pd.qcut(ranked, q=3, labels=["sparse", "medium", "dense"])
assigned = group[["dataset_id", "query_id"]].copy()
assigned["support_bucket"] = bins.astype(str)
rows.append(assigned)
return pd.concat(rows, ignore_index=True)
def _build_strength_tables() -> tuple[pd.DataFrame, pd.DataFrame, pd.DataFrame, pd.DataFrame]:
model_summary = pd.read_csv(BASE / "final" / "model_summary__v2.csv")
audit = pd.read_csv(RUN_DIR / "conditional_support_method_audit.csv")
bucket_map = _assign_primary_support_buckets(audit)
overall = model_summary[
[
"model_label",
"dataset_count",
"dependency_strength_similarity__mean",
"dependency_strength_similarity__ci95_low",
"dependency_strength_similarity__ci95_high",
"dependency_strength_similarity__ci95_radius",
]
].rename(
columns={
"dependency_strength_similarity__mean": "strength_mean",
"dependency_strength_similarity__ci95_low": "strength_ci95_low",
"dependency_strength_similarity__ci95_high": "strength_ci95_high",
"dependency_strength_similarity__ci95_radius": "strength_ci95_radius",
}
)
overall = overall.sort_values("strength_mean", ascending=False).reset_index(drop=True)
strength_rows = audit[audit["subitem_label"] == "Dependency strength similarity"].copy()
strength_rows = strength_rows.merge(bucket_map, on=["dataset_id", "query_id"], how="inner")
panel_strength = (
strength_rows.groupby(
["dataset_id", "dataset_prefix", "model_label", "support_bucket"],
as_index=False,
)["query_score"]
.mean()
.rename(columns={"query_score": "panel_strength"})
)
bucket_summary = (
panel_strength.groupby("support_bucket", as_index=False)
.agg(
strength_mean=("panel_strength", "mean"),
strength_std=("panel_strength", "std"),
panel_count=("panel_strength", "count"),
)
.reset_index(drop=True)
)
bucket_summary["strength_se"] = bucket_summary["strength_std"] / np.sqrt(bucket_summary["panel_count"])
bucket_summary["strength_ci95_radius"] = 1.96 * bucket_summary["strength_se"]
bucket_summary["bucket_label"] = bucket_summary["support_bucket"].map(SUPPORT_BUCKET_LABELS)
bucket_summary["support_bucket"] = pd.Categorical(
bucket_summary["support_bucket"], SUPPORT_BUCKET_ORDER, ordered=True
)
bucket_summary = bucket_summary.sort_values("support_bucket").reset_index(drop=True)
model_bucket = (
panel_strength.groupby(["model_label", "support_bucket"], as_index=False)["panel_strength"]
.mean()
.rename(columns={"panel_strength": "strength_mean"})
)
model_bucket["bucket_label"] = model_bucket["support_bucket"].map(SUPPORT_BUCKET_LABELS)
pivot = model_bucket.pivot(index="model_label", columns="support_bucket", values="strength_mean").reset_index()
for bucket in SUPPORT_BUCKET_ORDER:
if bucket not in pivot.columns:
pivot[bucket] = np.nan
pivot["range"] = pivot[SUPPORT_BUCKET_ORDER].max(axis=1) - pivot[SUPPORT_BUCKET_ORDER].min(axis=1)
pivot = pivot.sort_values("dense", ascending=False).reset_index(drop=True)
return overall, bucket_summary, model_bucket, pivot
def _plot_overall_strength(overall: pd.DataFrame) -> None:
fig, ax = plt.subplots(figsize=(11, 6))
x = np.arange(len(overall))
colors = [MODEL_COLORS.get(model, "#999999") for model in overall["model_label"]]
ax.bar(x, overall["strength_mean"], color=colors, edgecolor="black", linewidth=0.5)
ax.errorbar(
x,
overall["strength_mean"],
yerr=overall["strength_ci95_radius"],
fmt="none",
ecolor="black",
elinewidth=1,
capsize=3,
)
ax.set_xticks(x)
ax.set_xticklabels(overall["model_label"], rotation=45, ha="right")
ax.set_ylabel("Dependency strength similarity")
ax.set_title("Overall conditional strength by model")
ax.set_ylim(0, 0.8)
ax.grid(axis="y", alpha=0.25)
fig.tight_layout()
fig.savefig(OUT_DIR / "fig_strength_overall_model_bars.png", dpi=220)
fig.savefig(OUT_DIR / "fig_strength_overall_model_bars.pdf")
plt.close(fig)
def _plot_strength_bucket_summary(bucket_summary: pd.DataFrame) -> None:
fig, ax = plt.subplots(figsize=(7, 5))
x = np.arange(len(bucket_summary))
colors = [SUPPORT_BUCKET_COLORS[b] for b in bucket_summary["support_bucket"].astype(str)]
ax.bar(x, bucket_summary["strength_mean"], color=colors, edgecolor="black", linewidth=0.6)
ax.errorbar(
x,
bucket_summary["strength_mean"],
yerr=bucket_summary["strength_ci95_radius"],
fmt="none",
ecolor="black",
elinewidth=1,
capsize=4,
)
ax.set_xticks(x)
ax.set_xticklabels(bucket_summary["bucket_label"])
ax.set_ylabel("Dependency strength similarity")
ax.set_title("Conditional strength by support bucket")
ax.set_ylim(0, 0.45)
ax.grid(axis="y", alpha=0.25)
fig.tight_layout()
fig.savefig(OUT_DIR / "fig_strength_support_bucket_summary.png", dpi=220)
fig.savefig(OUT_DIR / "fig_strength_support_bucket_summary.pdf")
plt.close(fig)
def _plot_strength_by_model_bucket(model_bucket_pivot: pd.DataFrame) -> None:
models = model_bucket_pivot["model_label"].tolist()
x = np.arange(len(models))
width = 0.23
fig, ax = plt.subplots(figsize=(12, 6))
for idx, bucket in enumerate(SUPPORT_BUCKET_ORDER):
offset = (idx - 1) * width
ax.bar(
x + offset,
model_bucket_pivot[bucket],
width=width,
label=SUPPORT_BUCKET_LABELS[bucket],
color=SUPPORT_BUCKET_COLORS[bucket],
edgecolor="black",
linewidth=0.4,
)
ax.set_xticks(x)
ax.set_xticklabels(models, rotation=45, ha="right")
ax.set_ylabel("Dependency strength similarity")
ax.set_title("Conditional strength by model and support bucket")
ax.set_ylim(0, 0.45)
ax.grid(axis="y", alpha=0.25)
ax.legend(frameon=False, ncol=3, loc="upper center")
fig.tight_layout()
fig.savefig(OUT_DIR / "fig_strength_support_by_model.png", dpi=220)
fig.savefig(OUT_DIR / "fig_strength_support_by_model.pdf")
plt.close(fig)
def _write_report(
overall: pd.DataFrame,
bucket_summary: pd.DataFrame,
model_bucket_pivot: pd.DataFrame,
) -> None:
top = overall.iloc[0]
bottom = overall.iloc[-1]
dense = bucket_summary.loc[bucket_summary["support_bucket"].astype(str) == "dense", "strength_mean"].iloc[0]
medium = bucket_summary.loc[bucket_summary["support_bucket"].astype(str) == "medium", "strength_mean"].iloc[0]
sparse = bucket_summary.loc[bucket_summary["support_bucket"].astype(str) == "sparse", "strength_mean"].iloc[0]
flat = model_bucket_pivot.sort_values("range").head(3)
volatile = model_bucket_pivot.sort_values("range", ascending=False).head(3)
report = f"""# Strength-Only Conditional Analysis
## Scope
- Focus metric: `dependency_strength_similarity`
- Overall source: `final/model_summary__v2.csv`
- Support-bucket source: `locality_support_diagnostics/runs/20260502_064421_conditional_locality_support`
- Primary support variant: `scalar_filtered_local`
## What this isolates
This bundle ignores `direction_consistency` and `slice_level_consistency` and asks only one question:
> When the real data contains a conditional relationship, does the synthetic data preserve how *strong* that relationship is?
In downstream terms, this matters for tasks such as:
- deciding which conditional signals look strongest and therefore most actionable
- ranking features, slices, or segments by how tightly they track an outcome
- screening for candidate interactions before deeper local analysis
If strength is distorted, a downstream analyst may still see the right columns and the right report shape, but mis-rank which relationships deserve attention.
## Main findings
1. Overall model spread is real but not huge: the top model is `{top['model_label']}` at `{top['strength_mean']:.3f}`, while the weakest is `{bottom['model_label']}` at `{bottom['strength_mean']:.3f}`.
2. In the primary scalar filtered-local subset, support buckets are close: dense=`{dense:.3f}`, medium=`{medium:.3f}`, sparse=`{sparse:.3f}`.
3. This means strength does **not** show a clean sparse-support penalty in the current main support diagnostic.
4. Several models are almost flat across support buckets, especially:
- `{flat.iloc[0]['model_label']}` range=`{flat.iloc[0]['range']:.3f}`
- `{flat.iloc[1]['model_label']}` range=`{flat.iloc[1]['range']:.3f}`
- `{flat.iloc[2]['model_label']}` range=`{flat.iloc[2]['range']:.3f}`
5. The most bucket-sensitive models still do not follow one universal direction:
- `{volatile.iloc[0]['model_label']}` range=`{volatile.iloc[0]['range']:.3f}`
- `{volatile.iloc[1]['model_label']}` range=`{volatile.iloc[1]['range']:.3f}`
- `{volatile.iloc[2]['model_label']}` range=`{volatile.iloc[2]['range']:.3f}`
## Downstream interpretation
- Broad implication: many generators change conditional-strength estimates less across dense/medium/sparse local slices than one might expect. The support size of the slice is therefore not the main explanation for strength distortion.
- Practical implication: if a downstream user relies on synthetic data to decide *which* conditional relationships are strongest, the bigger risk is model-specific calibration of strength, not simply sparse local support.
- Reading by model family:
- `RealTabFormer` is strong overall and also stable across buckets, which makes it the cleanest strength-preserving model in the current panel.
- `BayesNet`, `ARF`, and `CTGAN` are comparatively flat across buckets, suggesting that for these models the strength story is more about their overall calibration level than about sensitivity to sparse slices.
- `TVAE`, `TabbyFlow`, and `TabSyn` vary more by bucket, but even there the movement is not monotonic dense-to-sparse collapse.
## Applicability note
The primary `Dense / Medium / Sparse` analysis is **not** a universal conditional-family split.
It applies only to filtered-local templates whose support can be defined as a scalar real row count:
- `Filtered Median Numeric Slice`
- `Filtered Sum in Numeric Band`
`Filtered Two-Dimensional Group Count` is excluded from the main bucket claim because its natural support object is a per-cell count distribution rather than one scalar filtered-row count.
"""
(OUT_DIR / "strength_focus_report.md").write_text(report, encoding="utf-8")
readme = """# Strength Focus
Strength-only conditional analysis artifacts.
Files:
- `overall_strength_by_model.csv`
- `strength_support_bucket_summary.csv`
- `strength_support_by_model.csv`
- `fig_strength_overall_model_bars.png`
- `fig_strength_support_bucket_summary.png`
- `fig_strength_support_by_model.png`
- `strength_focus_report.md`
Generate / refresh:
```bash
python Evaluation/query_fivepart_breakdown/conditional_breakdown/strength_focus/generate_strength_focus.py
```
"""
(OUT_DIR / "README.md").write_text(readme, encoding="utf-8")
def main() -> None:
OUT_DIR.mkdir(parents=True, exist_ok=True)
overall, bucket_summary, model_bucket, model_bucket_pivot = _build_strength_tables()
overall.to_csv(OUT_DIR / "overall_strength_by_model.csv", index=False)
bucket_summary.to_csv(OUT_DIR / "strength_support_bucket_summary.csv", index=False)
model_bucket_pivot.to_csv(OUT_DIR / "strength_support_by_model.csv", index=False)
model_bucket.to_csv(OUT_DIR / "strength_support_by_model_long.csv", index=False)
_plot_overall_strength(overall)
_plot_strength_bucket_summary(bucket_summary)
_plot_strength_by_model_bucket(model_bucket_pivot)
_write_report(overall, bucket_summary, model_bucket_pivot)
if __name__ == "__main__":
main()
|