code/scripts/11_rescore.py

"""Track 1. Add the new features to model-4's existing scores, gate on a clean val, ship.

    # measure (fit on g_stack, read g_val)
    python scripts/11_rescore.py --mode train
    # apply, only if the gate passed
    python scripts/11_rescore.py --mode test --bundle work/rescore/bundle.joblib

WHAT IS BEING TESTED. model-4 has no feature for "how many S1 share this candidate's name".
On its own stack, among empty-address candidates whose name_core equals the S1's:

    unique among the country's S1    mean p 0.714   actual match rate 0.982
    shared by 3 or more S1          mean p 0.280   actual 0.170

The model under-predicts the first group by 0.27 and over-predicts the third. Per coarse
group its calibration is perfect, so nothing short of conditioning on the competitor count
reveals it. These features supply exactly that count, plus the address-equivalence signal for
France, and the stacker is refit on top of model-4's own p1 and cross-encoder score.

THE ABLATION IS THE POINT. Four nested feature sets are fitted and read on g_val, so the gain
from each block is separable and a block that does nothing is visible rather than absorbed.

THE GATE. Ship only if the full set beats the baseline set by at least --min-gain on g_val.
Both numbers come from the same stacker on the same split, so the comparison is like for like
even if this stacker is not tuned as finely as model-4's own.
"""
from __future__ import annotations

import argparse
import json
import os
import sys
import time

import joblib
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 france_geo as G  # 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"]
M4 = os.environ.get("BERX_M4", "/scratch/user4/dai/model-4/work")


def prep_dir(split: str) -> str:
    return os.path.dirname(P.work("prep", split, "_"))


def add_all(df: pd.DataFrame, split: str, use_geo: bool = True):
    """Attach structure and geo features. `split` selects which source files to read."""
    d = prep_dir(split)
    s1 = pd.read_parquet(os.path.join(d, "source1.parquet"))
    pool = pd.read_parquet(os.path.join(d, "pool.parquet"))

    # ---- the six measured name-sharing features (MODEL5_FALSE_NEGATIVE_FIXES.md s3.1)
    t0 = time.time()
    print(f"  name_core / name_ns over {len(s1):,} S1 and {len(pool):,} pool rows ...",
          flush=True)
    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()
    print(f"  name-sharing features: {time.time()-t0:.0f}s", flush=True)

    t0 = time.time()
    print(f"  building name index over {len(s1):,} S1 rows ...", flush=True)
    ni = S.build_name_index(s1)
    df = S.add_features(df, s1["business_name"].to_numpy(), s1["country"].to_numpy(),
                        pool["business_name"].to_numpy(), pool["business_address"].to_numpy(),
                        pool["country"].to_numpy(), ni["count"], pcol="ce")
    print(f"  structure features: {time.time()-t0:.0f}s")
    feats = list(BASE) + S.NAME_FEATURES + S.FEATURE_NAMES

    gpath = P.work("geo", f"equiv_{split}.json")
    if use_geo and os.path.exists(gpath):
        t0 = time.time()
        tables = json.load(open(gpath, encoding="utf-8"))
        matchers = {c: G.GeoMatcher(t) for c, t in tables.items()}
        qcn = s1["country"].to_numpy()
        qi, ci = df["q"].to_numpy(), df["c"].to_numpy()
        # Parse only the addresses these candidates reference.
        uq, uc = np.unique(qi), np.unique(ci)
        qa, pa = s1["business_address"].to_numpy(), pool["business_address"].to_numpy()
        qp = {int(i): G.parse_address(qa[i]) for i in uq}
        pp = {int(i): G.parse_address(pa[i]) for i in uc}
        blank = {k: 0.0 for k in G.FEATURE_NAMES}
        cols = {k: np.zeros(len(df), np.float32) for k in G.FEATURE_NAMES}
        stepj = max(1, len(df) // 10)
        for j in range(len(df)):
            if j % stepj == 0 and j:
                print(f"    geo {j:>12,}/{len(df):,} ({j/len(df):5.1%})", flush=True)
            m = matchers.get(qcn[qi[j]])
            r = m.features_pre(qp[int(qi[j])], pp[int(ci[j])]) if m is not None else blank
            for k in G.FEATURE_NAMES:
                cols[k][j] = r[k]
        for k in G.FEATURE_NAMES:
            df[k] = cols[k]
        feats += G.FEATURE_NAMES
        print(f"  geo features: {time.time()-t0:.0f}s")
    elif use_geo:
        print(f"  no geo table at {gpath}; continuing without it")
    return df, feats


def truths_for(qs, gt_q, gt_c):
    keep = np.isin(gt_q, qs)
    out = {int(q): set() for q in qs}
    for q, c in zip(gt_q[keep], gt_c[keep]):
        out[int(q)].add(int(c))
    return out


def evaluate_block(tr, va, feats, truths, ids, name, n_folds=3):
    b = F.fit_stacker(tr, feats, label="y", group="q", n_folds=n_folds)
    va = va.copy()
    va["p"] = F.apply_stacker(b, va)
    # Both decision rules, because model-4 selects by expected-F0.5 and a flat threshold is
    # a different rule. Reporting only the winner would hide whether a gain came from the
    # features or from swapping the rule underneath them.
    best, f_thr = 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"  {name:<26} {len(feats):>3} feats   {best['method']:<9} "
          f"F0.5 {best['f05']:.5f}  (flat thr {f_thr:.5f})   "
          f"P {m['precision']:.4f}  R {m['recall']:.4f}")
    b["threshold"] = best["threshold"]
    b["method"] = best["method"]
    b["miss_mass"] = best["miss_mass"]
    return b, best["f05"]


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--mode", required=True, choices=["train", "test"])
    ap.add_argument("--bundle", default="")
    ap.add_argument("--min-gain", type=float, default=0.0015,
                    help="required g_val gain of the full set over the baseline set")
    ap.add_argument("--no-geo", action="store_true")
    ap.add_argument("--out", default="")
    a = ap.parse_args()

    if a.mode == "train":
        df = pd.read_parquet(P.work("m4", "train.parquet"))
        df, feats = add_all(df, "train", use_geo=not a.no_geo)
        tr = df[df.grp == "stack"]
        va = df[df.grp == "val"]
        t = np.load(P.work("m4", "truth.npz"))
        ids = sorted(set(va["q"].unique().tolist()))
        truths = truths_for(np.array(ids), t["gt_q"], t["gt_c"])
        nt = np.mean([len(v) for v in truths.values()])
        print(f"\nfit on {len(tr):,} stack rows / {tr.q.nunique():,} S1")
        print(f"read on {len(va):,} val rows / {len(ids):,} S1, mean true matches {nt:.2f}\n")

        # The six measured features get their own block, BEFORE my own set, so the number
        # that lands can be compared directly against the +0.0032 he measured. Putting them
        # last would have let my features absorb credit for them.
        blocks = [
            ("baseline (model-4 cols)", BASE),
            ("+ 6 name-sharing [measured]", BASE + S.NAME_FEATURES),
            ("+ my structure", BASE + S.NAME_FEATURES
             + [f for f in S.FEATURE_NAMES if f in df.columns]),
            ("+ geo", feats),
        ]
        results = []
        for name, fs in blocks:
            fs = [f for f in fs if f in df.columns]
            results.append((name, *evaluate_block(tr, va, fs, truths, ids, name)))

        base_f05 = results[0][2]
        best_name, best_b, best_f05 = max(results, key=lambda r: r[2])
        gain = best_f05 - base_f05
        print(f"\n  baseline {base_f05:.5f}  ->  best '{best_name}' {best_f05:.5f}  "
              f"({gain:+.5f})")
        print(f"  model-4's own val F0.5 for reference: 0.98298")

        p = a.bundle or P.work("rescore", "bundle.joblib")
        joblib.dump(best_b, p)
        json.dump({"base": base_f05, "best": best_f05, "gain": gain,
                   "threshold": best_b["threshold"], "block": best_name,
                   "pass": bool(gain >= a.min_gain)},
                  open(P.work("rescore", "gate.json"), "w"), indent=1)

        if gain >= a.min_gain:
            print(f"\n  GATE A PASS ({gain:+.5f} >= {a.min_gain}). Run --mode test.")
        else:
            print(f"\n  GATE A FAIL ({gain:+.5f} < {a.min_gain}).")
            print("  The global information is already in model-4's scores. Do not ship this;")
            print("  keep the existing submission and put the time into retrieval instead.")
            sys.exit(3)
        return

    assert a.bundle, "--mode test needs --bundle"
    b = joblib.load(a.bundle)
    df = pd.read_parquet(P.work("m4", "test.parquet"))
    df, _ = add_all(df, "test", use_geo=not a.no_geo)
    missing = [f for f in b["feats"] if f not in df.columns]
    assert not missing, f"test table is missing features the bundle needs: {missing}"
    df["p"] = F.apply_stacker(b, df)

    d = prep_dir("test")
    s1 = pd.read_parquet(os.path.join(d, "source1.parquet"))
    pool = pd.read_parquet(os.path.join(d, "pool.parquet"))
    pred = F.predict_sets(df, b["threshold"], all_q=np.arange(len(s1)),
                          method=b.get("method", "threshold"),
                          miss_mass=b.get("miss_mass", 0.0))

    out = a.out or P.work("output_v2", "matching_results.tsv")
    os.makedirs(os.path.dirname(out), exist_ok=True)
    qid, pid = s1["entity_id"].to_numpy(), pool["entity_id"].to_numpy()
    with open(out, "w", encoding="utf-8", newline="") as fh:
        fh.write("source1_entity_id\tmatched_entity_ids\n")
        for i in range(len(qid)):
            fh.write(f"{qid[i]}\t{','.join(pid[c] for c in pred.get(i, []))}\n")
    npred = sum(len(v) for v in pred.values())
    nempty = sum(1 for v in pred.values() if not v)
    print(f"\nwrote {out}")
    print(f"  {len(qid):,} S1 rows   {npred:,} pairs   {nempty:,} empty ({nempty/len(qid):.2%})")
    print(f"  rule {b.get('method','threshold')}  threshold {b['threshold']:.3f}  "
          f"miss_mass {b.get('miss_mass',0.0)}")


if __name__ == "__main__":
    main()