code/src/metric.py

"""Challenge metric for AMLC 2026 business entity resolution.

Per S1 entity, with prediction set S and truth set T:

    T empty,  S empty      -> 1.0
    T empty,  S non-empty  -> 0.0
    T non-empty, S empty   -> 0.0
    otherwise P = |S&T|/|S|, R = |S&T|/|T|, F = 1.25*P*R / (0.25*P + R),
              and 0.0 when |S&T| = 0.

The leaderboard score is the macro average of F over all S1 entities in the
truth file (a missing prediction counts as the empty set).

Note: with tp = |S&T|, k = |S|, n = |T| the formula reduces to
F = 5*tp / (4*k + n), which is what src/decode.py optimises.

CLI:
    python -m src.metric --pred matching_results.tsv --truth train_ground_truth.tsv \
        [--groups s1.tsv [--group-col country]]
"""
from __future__ import annotations

import argparse
import csv
import sys
from typing import Iterable, Mapping

import numpy as np
import pandas as pd

__all__ = [
    "f05_entity",
    "entity_scores",
    "macro_f05",
    "read_id_list_tsv",
    "write_id_list_tsv",
    "read_groups",
    "breakdown",
]

ID_COL = "source1_entity_id"
LIST_COLS = ("matched_entity_ids", "candidate_entity_ids")

_EMPTY: frozenset = frozenset()


# ---------------------------------------------------------------------------
# Per-entity score
# ---------------------------------------------------------------------------
def f05_entity(pred: set, true: set) -> float:
    """F0.5 for one S1 entity, with the exact challenge edge rules."""
    if not true:
        return 0.0 if pred else 1.0
    if not pred:
        return 0.0
    if not isinstance(pred, (set, frozenset)):
        pred = set(pred)
    if not isinstance(true, (set, frozenset)):
        true = set(true)
    tp = len(pred & true)
    if tp == 0:
        return 0.0
    p = tp / len(pred)
    r = tp / len(true)
    return 1.25 * p * r / (0.25 * p + r)


def _prf(pred: set, true: set) -> tuple[float, float, float, int]:
    """(F0.5, precision or nan, recall or nan, tp) for one entity.

    Precision is nan when pred is empty, recall is nan when true is empty.
    """
    f = f05_entity(pred, true)
    if pred and true:
        if isinstance(pred, (set, frozenset)) and isinstance(true, (set, frozenset)):
            tp = len(pred & true)
        else:
            tp = len(set(pred) & set(true))
    else:
        tp = 0
    prec = tp / len(pred) if pred else np.nan
    rec = tp / len(true) if true else np.nan
    return f, prec, rec, tp


def entity_scores(pred: Mapping[str, set], truth: Mapping[str, set]) -> pd.DataFrame:
    """One row per truth key: f05, precision, recall, tp, n_pred, n_true.

    Keys missing from ``pred`` are scored as empty predictions; keys present
    only in ``pred`` are ignored (the score is defined over the truth file).
    """
    keys = list(truth.keys())
    n = len(keys)
    f = np.empty(n)
    prec = np.empty(n)
    rec = np.empty(n)
    tp = np.empty(n, dtype=np.int32)
    n_pred = np.empty(n, dtype=np.int32)
    n_true = np.empty(n, dtype=np.int32)
    get = pred.get
    for i, key in enumerate(keys):
        t = truth[key]
        s = get(key, _EMPTY)
        f[i], prec[i], rec[i], tp[i] = _prf(s, t)
        n_pred[i] = len(s)
        n_true[i] = len(t)
    return pd.DataFrame(
        {"f05": f, "precision": prec, "recall": rec, "tp": tp, "n_pred": n_pred, "n_true": n_true},
        index=pd.Index(keys, name=ID_COL),
    )


def macro_f05(pred: Mapping[str, set], truth: Mapping[str, set]) -> float:
    """Macro F0.5 averaged over ALL keys of ``truth`` (missing pred = empty)."""
    if not truth:
        return float("nan")
    get = pred.get
    total = 0.0
    for key, t in truth.items():
        total += f05_entity(get(key, _EMPTY), t)
    return total / len(truth)


# ---------------------------------------------------------------------------
# I/O
# ---------------------------------------------------------------------------
def _clean(tok: str) -> str:
    tok = tok.strip()
    if len(tok) >= 2 and tok[0] == tok[-1] and tok[0] in "\"'":
        tok = tok[1:-1].strip()
    return tok


def _parse_list(field: str) -> set:
    field = _clean(field)
    if not field:
        return set()
    if field[0] == "[" and field[-1] == "]":  # tolerate python-list style
        field = field[1:-1]
    out = set()
    for tok in field.split(","):
        tok = _clean(tok)
        if tok:
            out.add(sys.intern(tok))
    return out


def read_id_list_tsv(path) -> dict[str, set]:
    """Read a challenge TSV ``source1_entity_id<TAB>matched_entity_ids``.

    The header's second column may also be ``candidate_entity_ids``. The second
    column is a comma-separated list and may be empty or missing. Streams line
    by line; repeated S1 ids are unioned. Candidate ids are interned so that ids
    shared across rows are stored once.
    """
    out: dict[str, set] = {}
    with open(path, "r", encoding="utf-8", newline="") as fh:
        first = fh.readline()
        if not first:
            return out
        head = [_clean(c).lower() for c in first.rstrip("\r\n").split("\t")]
        is_header = head[0] == ID_COL or (len(head) > 1 and head[1] in LIST_COLS)
        lines: Iterable[str] = fh
        if not is_header:
            lines = _chain_first(first, fh)
        for line in lines:
            line = line.rstrip("\r\n")
            if not line.strip():
                continue
            key, sep, rest = line.partition("\t")
            key = _clean(key)
            if not key:
                continue
            vals = _parse_list(rest) if sep else set()
            cur = out.get(key)
            if cur is None:
                out[key] = vals
            else:
                cur |= vals
    return out


def _chain_first(first: str, fh):
    yield first
    yield from fh


def write_id_list_tsv(pred: Mapping[str, Iterable[str]], path, all_ids: Iterable[str] | None = None,
                      list_col: str = "matched_entity_ids") -> None:
    """Write ``pred`` in the challenge format. ``all_ids`` (if given) fixes the
    row set/order; ids with no prediction get an empty second column."""
    keys = list(all_ids) if all_ids is not None else list(pred.keys())
    get = pred.get
    with open(path, "w", encoding="utf-8", newline="") as fh:
        fh.write(f"{ID_COL}\t{list_col}\n")
        for key in keys:
            vals = get(key) or ()
            fh.write(f"{key}\t{','.join(map(str, vals))}\n")


def read_groups(path, group_col: str = "country", id_col: str | None = None) -> dict[str, str]:
    """Map S1 id -> group label from a TSV/CSV (e.g. the S1 entity file).

    Only the id and group columns are loaded. The id column defaults to
    ``source1_entity_id``, then ``entity_id``, then the first column.
    """
    with open(path, "r", encoding="utf-8") as fh:
        header = fh.readline().rstrip("\r\n")
    sep = "\t" if "\t" in header else ","
    cols = [_clean(c) for c in header.split(sep)]
    if id_col is None:
        lower = [c.lower() for c in cols]
        for cand in (ID_COL, "entity_id", "id"):
            if cand in lower:
                id_col = cols[lower.index(cand)]
                break
        else:
            id_col = cols[0]
    if group_col not in cols:
        raise ValueError(f"group column {group_col!r} not in {path}; columns: {cols}")
    # TSV: no quote processing, so stray quotes in free-text columns (names,
    # addresses) cannot swallow lines; ids/labels are unquoted afterwards.
    quoting = csv.QUOTE_NONE if sep == "\t" else csv.QUOTE_MINIMAL
    df = pd.read_csv(path, sep=sep, usecols=[id_col, group_col], dtype=str,
                     keep_default_na=False, quoting=quoting)
    ids = [_clean(x) for x in df[id_col].tolist()]
    labs = [_clean(x) for x in df[group_col].tolist()]
    return dict(zip(ids, labs))


# ---------------------------------------------------------------------------
# Breakdown
# ---------------------------------------------------------------------------
_BUCKETS = ["0", "1", "2", "3", "4", "5+"]


def _summarise(g: pd.DataFrame, total_loss: float, n_total: int) -> dict:
    single = g["n_true"].to_numpy() == 0
    loss = float((1.0 - g["f05"]).sum())
    return {
        "n": len(g),
        "share": len(g) / n_total if n_total else np.nan,
        "macro_f05": float(g["f05"].mean()) if len(g) else np.nan,
        # precision averaged over entities with a non-empty prediction,
        # recall averaged over entities with a non-empty truth set
        "mean_precision": float(np.nanmean(g["precision"])) if g["precision"].notna().any() else np.nan,
        "mean_recall": float(np.nanmean(g["recall"])) if g["recall"].notna().any() else np.nan,
        "n_singleton": int(single.sum()),
        # fraction of truth-empty entities predicted empty
        "singleton_acc": float((g["n_pred"].to_numpy()[single] == 0).mean()) if single.any() else np.nan,
        "pred_empty_rate": float((g["n_pred"] == 0).mean()) if len(g) else np.nan,
        "mean_n_pred": float(g["n_pred"].mean()) if len(g) else np.nan,
        "mean_n_true": float(g["n_true"].mean()) if len(g) else np.nan,
        # share of the total (1 - F) loss that this slice accounts for
        "loss_share": loss / total_loss if total_loss > 0 else 0.0,
    }


def breakdown(pred: Mapping[str, set], truth: Mapping[str, set],
              groups: Mapping[str, str] | None = None,
              scores: pd.DataFrame | None = None) -> pd.DataFrame:
    """Per-slice diagnostics.

    Rows: ``overall``; one per group label (if ``groups`` is given; S1 ids not in
    ``groups`` go to ``<missing>``); one per true-match-count bucket
    (0,1,2,3,4,5+). Columns: n, share, macro_f05, mean_precision, mean_recall,
    n_singleton, singleton_acc, pred_empty_rate, mean_n_pred, mean_n_true,
    loss_share. The first two index levels are (dimension, slice).
    """
    s = entity_scores(pred, truth) if scores is None else scores
    n_total = len(s)
    total_loss = float((1.0 - s["f05"]).sum())
    rows: list[tuple[str, str, dict]] = [("overall", "all", _summarise(s, total_loss, n_total))]

    if groups is not None:
        glab = pd.Series([groups.get(k, "<missing>") for k in s.index], index=s.index)
        order = glab.value_counts().index
        for lab in order:
            rows.append(("group", str(lab), _summarise(s[glab.to_numpy() == lab], total_loss, n_total)))

    bucket = np.minimum(s["n_true"].to_numpy(), 5)
    for b in range(6):
        sub = s[bucket == b]
        if len(sub):
            rows.append(("true_count", _BUCKETS[b], _summarise(sub, total_loss, n_total)))

    idx = pd.MultiIndex.from_tuples([(d, v) for d, v, _ in rows], names=["dimension", "slice"])
    return pd.DataFrame([r for _, _, r in rows], index=idx)


# ---------------------------------------------------------------------------
# CLI
# ---------------------------------------------------------------------------
def main(argv=None) -> int:
    ap = argparse.ArgumentParser(description="Macro F0.5 for AMLC 2026 entity resolution")
    ap.add_argument("--pred", required=True, help="matching_results.tsv")
    ap.add_argument("--truth", required=True, help="ground-truth TSV in the same format")
    ap.add_argument("--groups", default=None, help="TSV/CSV with S1 id + group column (e.g. s1.tsv)")
    ap.add_argument("--group-col", default="country")
    ap.add_argument("--id-col", default=None, help="id column in --groups file (auto-detected)")
    args = ap.parse_args(argv)

    truth = read_id_list_tsv(args.truth)
    pred = read_id_list_tsv(args.pred)
    extra = sum(1 for k in pred if k not in truth)
    missing = sum(1 for k in truth if k not in pred)
    groups = read_groups(args.groups, args.group_col, args.id_col) if args.groups else None

    scores = entity_scores(pred, truth)
    print(f"truth entities: {len(truth):,}   pred rows: {len(pred):,}   "
          f"truth ids missing from pred (scored as empty): {missing:,}   "
          f"pred ids not in truth (ignored): {extra:,}")
    print(f"macro F0.5 = {scores['f05'].mean():.6f}")
    bd = breakdown(pred, truth, groups, scores=scores)
    with pd.option_context("display.max_rows", 500, "display.max_columns", None, "display.width", 250,
                           "display.float_format", "{:.4f}".format):
        print(bd)
    return 0


if __name__ == "__main__":
    raise SystemExit(main())