File size: 8,618 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
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
"""Diagnose the false positives. 0.0041 of val loss, 24% of the total, undiagnosed by anyone.

Every document in this project has attacked false NEGATIVES. FALSE_NEGATIVES_2026-09-27.md
s1 lists FP at 0.0041 and MODEL5_FALSE_NEGATIVE_FIXES.md s6 records it as "not diagnosed
yet [?]". It is now the largest unexamined bucket in the US/India half, and the arithmetic
says US/India must gain +0.0045 val for the leaderboard to reach 0.990 - so this bucket
alone is roughly the size of what is needed.

THE QUESTION THAT DECIDES WHAT TO BUILD. A predicted pair (q, c) that is wrong is wrong in
one of three ways, and they need completely different fixes:

  A. c belongs to a DIFFERENT S1, and that S1 is also in our candidate table.
     -> an arbitration / ranking failure. We had the right answer and chose the wrong one.
     -> fixable with better global assignment. No new retrieval, no new scoring.

  B. c belongs to a different S1 that is NOT in our candidate table for c.
     -> a retrieval failure wearing a false-positive costume: the true owner never competed.
     -> fixable only by retrieving more.

  C. c belongs to NO S1 (a true distractor).
     -> a scoring failure. 26.8% of addressed pool records are distractors, and 22.5% of
        those deliberately reuse a real S1 name.
     -> fixable only with a better scorer or better features.

The split between A, B and C is the whole answer, and one pass over the labelled val split
settles it. Run:

    python scripts/20_fp_diag.py --bundle work/rescore/bundle.joblib

With no bundle it fits a quick stacker on g_stack first, so it can run standalone.
"""
from __future__ import annotations

import argparse
import collections
import os
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 evaluate as E  # noqa: E402
from berx import fuse as F  # noqa: E402
from berx import paths as P  # noqa: E402
from berx import structure as S  # noqa: E402

BASE = ["p1", "p1_rank", "p1_gap", "ce"]


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--bundle", default="")
    ap.add_argument("--split", default="train")
    ap.add_argument("--max-examples", type=int, default=12)
    a = ap.parse_args()

    df = pd.read_parquet(P.work("m4", "train.parquet"))
    d = os.path.dirname(P.work("prep", a.split, "_"))
    s1 = pd.read_parquet(os.path.join(d, "source1.parquet"))
    pool = pd.read_parquet(os.path.join(d, "pool.parquet"))

    # the six measured features, so the decision resembles what actually ships
    s1 = pd.concat([s1, S.name_key_columns(s1)], axis=1)
    pool = pd.concat([pool, S.name_key_columns(pool)], axis=1)
    pool["addr_empty"] = pool["business_address"].astype(str).str.strip() == ""
    nf = S.name_features(df, s1, pool)
    for c in S.NAME_FEATURES:
        df[c] = nf[c].to_numpy()
    feats = BASE + S.NAME_FEATURES

    tr, va = df[df.grp == "stack"], df[df.grp == "val"].copy()
    if a.bundle and os.path.exists(a.bundle):
        import joblib
        b = joblib.load(a.bundle)
        feats = [f for f in b["feats"] if f in va.columns]
        if len(feats) != len(b["feats"]):
            print("  bundle wants features this table lacks; refitting instead")
            b = F.fit_stacker(tr, BASE + S.NAME_FEATURES, label="y", group="q", n_folds=3)
    else:
        b = F.fit_stacker(tr, feats, label="y", group="q", n_folds=3)
    va["p"] = F.apply_stacker(b, va)

    t = np.load(P.work("m4", "truth.npz"))
    ids = sorted(va["q"].unique().tolist())
    idset = set(ids)
    truths = {q: set() for q in ids}
    keep = np.isin(t["gt_q"], ids)
    for q, c in zip(t["gt_q"][keep], t["gt_c"][keep]):
        truths[int(q)].add(int(c))

    # owner[c] -> the S1 that owns pool record c, over ALL ground truth, not just val
    owner = dict(zip(t["gt_c"].tolist(), t["gt_q"].tolist()))

    best, _ = F.tune_decision(va, truths, ids)
    pred = F.predict_sets(va, best["threshold"], all_q=np.asarray(ids),
                          method=best["method"], miss_mass=best["miss_mass"])
    m, _ = E.macro_f05(pred, truths, ids)
    print(f"decision: {best['method']}  F0.5 {m['f05']:.5f}  "
          f"P {m['precision']:.4f}  R {m['recall']:.4f}\n")

    fp = [(q, c) for q, cs in pred.items() for c in cs if c not in truths.get(q, ())]
    n_pred = sum(len(v) for v in pred.values())
    print(f"predicted {n_pred:,} pairs, {len(fp):,} false positives "
          f"({len(fp)/max(n_pred,1):.2%})\n")

    # who else was competing for each record, in OUR table
    claim = collections.defaultdict(list)
    for q, c, p in zip(va["q"].to_numpy(), va["c"].to_numpy(), va["p"].to_numpy()):
        claim[int(c)].append((int(q), float(p)))

    cat = collections.Counter()
    rank_of_true = collections.Counter()
    margins = []
    examples = {"A": [], "B": [], "C": []}
    for q, c in fp:
        o = owner.get(c)
        if o is None:
            cat["C distractor (belongs to no S1)"] += 1
            examples["C"].append((q, c, None))
            continue
        competitors = claim.get(c, [])
        has_true = any(qq == o for qq, _ in competitors)
        if has_true:
            cat["A true owner WAS in our table (we chose wrong)"] += 1
            ranked = sorted(competitors, key=lambda x: -x[1])
            pos = [i for i, (qq, _) in enumerate(ranked) if qq == o]
            rank_of_true[min(pos[0], 5) if pos else 99] += 1
            pw = dict(competitors)
            margins.append(pw[q] - pw[o])
            examples["A"].append((q, c, o))
        else:
            cat["B true owner NOT in our table (retrieval miss)"] += 1
            examples["B"].append((q, c, o))

    print("=" * 74)
    print("WHERE THE FALSE POSITIVES COME FROM")
    print("=" * 74)
    for k, v in cat.most_common():
        print(f"  {k:<48} {v:>7,}  {v/max(len(fp),1):6.1%}")

    if rank_of_true:
        print("\n  For group A, where did the TRUE owner rank among competitors?")
        tot = sum(rank_of_true.values())
        for r in sorted(rank_of_true):
            lab = f"rank {r}" if r < 5 else ("rank 5+" if r < 99 else "not ranked")
            print(f"    {lab:<12} {rank_of_true[r]:>7,}  {rank_of_true[r]/tot:6.1%}")
        mg = np.array(margins)
        print(f"\n  Score margin by which the WRONG S1 beat the true owner:")
        print(f"    median {np.median(mg):.4f}   p25 {np.percentile(mg,25):.4f}   "
              f"p75 {np.percentile(mg,75):.4f}")
        print(f"    within 0.05: {(mg < 0.05).mean():.1%}   within 0.15: {(mg < 0.15).mean():.1%}")
        print("    A small margin means the ordering is nearly right and a better")
        print("    tie-break could flip it. A large one means the scorer is confident")
        print("    and wrong, which needs features, not arbitration.")

    print("\n" + "=" * 74)
    print("WHAT THE FALSE POSITIVES LOOK LIKE")
    print("=" * 74)
    va_idx = {(int(q), int(c)): i for i, (q, c) in
              enumerate(zip(va["q"].to_numpy(), va["c"].to_numpy()))}
    ae = pool["addr_empty"].to_numpy()
    ctry = s1["country"].to_numpy()
    for grp in ("A", "B", "C"):
        ex = examples[grp]
        if not ex:
            continue
        qs = np.array([e[0] for e in ex])
        cs = np.array([e[1] for e in ex])
        print(f"\n  group {grp}: {len(ex):,}   empty-address {ae[cs].mean():.1%}   "
              f"by country {dict(collections.Counter(ctry[qs]).most_common())}")
        for q, c, o in ex[:a.max_examples // 3]:
            qn = s1['business_name'].iloc[q]
            cn = pool['business_name'].iloc[c]
            ca = pool['business_address'].iloc[c]
            line = f"    S1 {qn[:34]!r:38} -> {cn[:30]!r:34} | {str(ca)[:26]!r}"
            if o is not None:
                line += f"  TRUE OWNER {s1['business_name'].iloc[o][:26]!r}"
            print(line)

    print("\n" + "=" * 74)
    print("WHAT THIS MEANS")
    print("=" * 74)
    a_share = cat["A true owner WAS in our table (we chose wrong)"] / max(len(fp), 1)
    b_share = cat["B true owner NOT in our table (retrieval miss)"] / max(len(fp), 1)
    c_share = cat["C distractor (belongs to no S1)"] / max(len(fp), 1)
    print(f"  A {a_share:5.1%} fixable by better ASSIGNMENT  (no new retrieval or scoring)")
    print(f"  B {b_share:5.1%} needs better RETRIEVAL for the true owner")
    print(f"  C {c_share:5.1%} needs a better SCORER or decoy features")
    print(f"\n  The 0.0041 of val loss splits roughly "
          f"{0.0041*a_share:.4f} / {0.0041*b_share:.4f} / {0.0041*c_share:.4f}.")


if __name__ == "__main__":
    main()