File size: 7,304 Bytes
d231962 | 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 | """Step A. Mine administrative equivalences, and VALIDATE them where labels exist.
The validation is the whole reason this script is longer than the miner. France has no labels,
so a France-only change can never be checked directly - and the last France change made on
reasoning alone pointed the threshold in the wrong direction.
The honest substitute: the same generator produced US and India, the same swap appears there
(district vs state, county vs state), and those DO have labels. So the miner is run on all
three countries, and on US/India the mined features are scored against ground truth. If they
separate true pairs from false ones there, the France table is trustworthy for the same
reason. If they do not, the France table is discarded rather than shipped on faith.
Runs on CPU in a few minutes. No GPU, no model, no artifacts from any other pipeline.
"""
from __future__ import annotations
import argparse
import json
import os
import random
import sys
import numpy as np
import pandas as pd
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from berx import france_geo as G # noqa: E402
from berx import paths as P # noqa: E402
def load_prep(split: str):
d = os.path.dirname(P.work("prep", split, "_"))
s1 = pd.read_parquet(os.path.join(d, "source1.parquet"))
pool = pd.read_parquet(os.path.join(d, "pool.parquet"))
return s1, pool
def validate(s1: pd.DataFrame, pool: pd.DataFrame, gt_path: str, matcher: G.GeoMatcher,
country: str, n_pos: int = 4000, seed: int = 42):
"""Do the mined features separate true pairs from hard false ones, on a labelled country?
The negatives are deliberately HARD: for each sampled true pair, a different pool record
of the same country whose name shares a token with the S1. Random negatives would be
separated by the name alone and would tell us nothing about whether the geo features work.
"""
from berx import textnorm as T
gt = pd.read_csv(gt_path, sep="\t", dtype=str, keep_default_na=False, na_filter=False) \
if gt_path.endswith(".tsv") else pd.read_parquet(gt_path)
s1c = s1[s1["country"] == country]
s1_idx = dict(zip(s1c["entity_id"], range(len(s1c))))
s1a = dict(zip(s1c["entity_id"], s1c["business_address"]))
s1n = dict(zip(s1c["entity_id"], s1c["business_name"]))
poolc = pool[pool["country"] == country]
pa = dict(zip(poolc["entity_id"], poolc["business_address"]))
pn = dict(zip(poolc["entity_id"], poolc["business_name"]))
tok_index: dict[str, list] = {}
for eid, nm in pn.items():
for t in set(T.tokens(T.norm_text(nm))):
tok_index.setdefault(t, []).append(eid)
rng = random.Random(seed)
rows = []
gt = gt[gt["source1_entity_id"].isin(s1_idx)]
for _, r in gt.sample(min(n_pos, len(gt)), random_state=seed).iterrows():
q = r["source1_entity_id"]
mids = [x for x in str(r["matched_entity_ids"]).split(",") if x in pa]
if not mids:
continue
c = rng.choice(mids)
rows.append((s1a[q], pa[c], 1))
cand = []
for t in set(T.tokens(T.norm_text(s1n[q]))):
cand += tok_index.get(t, [])[:50]
cand = [x for x in cand if x not in set(mids)]
if cand:
rows.append((s1a[q], pa[rng.choice(cand)], 0))
if not rows:
return None
F = pd.DataFrame([{**matcher.features(a, b), "y": y} for a, b, y in rows])
rep = {}
for c in G.FEATURE_NAMES:
t, f = F[F.y == 1][c], F[F.y == 0][c]
rep[c] = {"true_mean": round(float(t.mean()), 4),
"false_mean": round(float(f.mean()), 4),
"separation": round(float(t.mean() - f.mean()), 4)}
# the headline check: does equiv beat exact at telling true from false?
rep["_n"] = len(F)
rep["_lift_equiv_over_exact"] = round(
rep["geo_equiv"]["separation"] - rep["geo_exact"]["separation"], 4)
return rep
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--split", default="test")
ap.add_argument("--validate-on", default="train",
help="split that has ground truth, for the US/India sanity check")
ap.add_argument("--min-support", type=int, default=8)
ap.add_argument("--min-ratio", type=float, default=0.35)
ap.add_argument("--skip-validate", action="store_true")
a = ap.parse_args()
s1, pool = load_prep(a.split)
tables = {}
for country in sorted(set(pool["country"])):
sub = pool[pool["country"] == country]["business_address"]
sub1 = s1[s1["country"] == country]["business_address"]
print(f"\n== {country}: mining over {len(sub) + len(sub1):,} addresses")
t = G.mine(list(sub) + list(sub1), min_support=a.min_support, min_ratio=a.min_ratio)
n_pairs = sum(len(v) for v in t["equiv"].values()) // 2
print(f" admin names {t['n_admin_names']:,} equivalences {n_pairs:,}")
ex = [(k, sorted(v.items(), key=lambda kv: -kv[1])[:3])
for k, v in list(t["equiv"].items())[:8]]
for k, v in ex:
print(f" {k!r:32} ~ {[(x, s) for x, s in v]}")
tables[country] = t
out = P.work("geo", f"equiv_{a.split}.json")
with open(out, "w", encoding="utf-8") as fh:
json.dump(tables, fh, ensure_ascii=False)
print(f"\nwrote {out}")
if a.skip_validate:
return
print("\n" + "=" * 72)
print("VALIDATION on labelled countries - does this help where we can check it?")
print("=" * 72)
s1t, poolt = load_prep(a.validate_on)
gt = P.ground_truth(a.validate_on)
report = {}
for country in ("US", "India"):
if country not in set(poolt["country"]):
continue
# validate with the table mined on the SAME split, or the check is contaminated by
# the test pool having been used to build the thing being tested
tt = G.mine(list(poolt[poolt["country"] == country]["business_address"]),
min_support=a.min_support, min_ratio=a.min_ratio)
rep = validate(s1t, poolt, gt, G.GeoMatcher(tt), country)
if rep:
report[country] = rep
print(f"\n {country} (n={rep['_n']:,})")
for c in G.FEATURE_NAMES:
r = rep[c]
print(f" {c:<16} true {r['true_mean']:>8.4f} false {r['false_mean']:>8.4f}"
f" sep {r['separation']:>+8.4f}")
print(f" -> equiv beats exact by {rep['_lift_equiv_over_exact']:+.4f}")
ok = all(r["_lift_equiv_over_exact"] > 0.002 for r in report.values()) and report
print("\n" + ("PASS - the equivalence table adds separation on labelled countries, so the"
"\n France table is worth shipping as features."
if ok else
"FAIL - equiv does not beat exact where we can check it. Do NOT ship the"
"\n France table; the swap is either rarer or differently shaped than"
"\n assumed, and shipping it on reasoning alone is how the last France"
"\n change went in backwards."))
with open(P.work("geo", "validation.json"), "w") as fh:
json.dump(report, fh, indent=1)
if __name__ == "__main__":
main()
|