code/rich_randomwalk_stack.py

"""Add rich content features to the systematic random-walk stack."""

from __future__ import annotations

import argparse
import importlib.util
import sys
from pathlib import Path

import lightgbm as lgb
import numpy as np
import pandas as pd
from gensim.models import Word2Vec
from sklearn.metrics import precision_recall_curve, roc_auc_score
from sklearn.model_selection import StratifiedKFold


def load_module(name: str, path: Path):
    spec = importlib.util.spec_from_file_location(name, path)
    module = importlib.util.module_from_spec(spec)
    assert spec.loader is not None
    sys.modules[name] = module
    spec.loader.exec_module(module)
    return module


def best_f1(y: np.ndarray, s: np.ndarray):
    p, r, t = precision_recall_curve(y, s)
    f = 2 * p * r / (p + r + 1e-12)
    i = int(np.argmax(f))
    th = float(t[i]) if i < len(t) else 0.5
    return float(f[i]), th, float(roc_auc_score(y, s)), float(p[i]), float(r[i])


def fit_lgb_oof(X: np.ndarray, y: np.ndarray, seed: int, n_splits: int, leaves: int = 31) -> np.ndarray:
    oof = np.zeros(len(y), dtype=np.float32)
    skf = StratifiedKFold(n_splits=n_splits, shuffle=True, random_state=seed)
    for fold, (tr, va) in enumerate(skf.split(X, y), start=1):
        clf = lgb.LGBMClassifier(
            n_estimators=1200,
            learning_rate=0.025,
            num_leaves=leaves,
            subsample=0.9,
            colsample_bytree=0.9,
            reg_lambda=7.0,
            min_child_samples=80,
            objective="binary",
            n_jobs=8,
            verbose=-1,
            random_state=seed + fold,
        )
        clf.fit(X[tr], y[tr])
        oof[va] = clf.predict_proba(X[va])[:, 1].astype(np.float32)
    return oof


def fit_full_predict(X: np.ndarray, y: np.ndarray, Xt: np.ndarray, seed: int, leaves: int = 31) -> np.ndarray:
    clf = lgb.LGBMClassifier(
        n_estimators=1400,
        learning_rate=0.022,
        num_leaves=leaves,
        subsample=0.9,
        colsample_bytree=0.9,
        reg_lambda=7.0,
        min_child_samples=80,
        objective="binary",
        n_jobs=8,
        verbose=-1,
        random_state=seed,
    )
    clf.fit(X, y)
    return clf.predict_proba(Xt)[:, 1].astype(np.float32)


def write_sub(path: Path, score: np.ndarray, known: np.ndarray, anchor: np.ndarray, ratio: float) -> tuple[float, int]:
    pred = np.zeros(len(score), dtype=np.int8)
    pred[np.argsort(score, kind="mergesort")[-int(round(len(score) * ratio)):]] = 1
    pred[known] = 1
    pd.DataFrame({"Index": np.arange(len(pred), dtype=np.int64), "Predicted": pred}).to_csv(path, index=False)
    return float(pred.mean()), int((pred != anchor).sum())


def main() -> None:
    ap = argparse.ArgumentParser()
    ap.add_argument("--package-root", type=Path, default=Path(__file__).resolve().parents[1])
    ap.add_argument("--split-seed", type=int, default=202)
    ap.add_argument("--main-val-score-file", type=Path, default=None)
    ap.add_argument("--seed", type=int, default=202)
    ap.add_argument("--n-splits", type=int, default=5)
    args = ap.parse_args()

    root = args.package_root.resolve()
    main_val = args.main_val_score_file or root / "validation_runs/dynamic_seed202/dyn202_l2d512_bpr_bigbatch_more/scores/val_vanilla_ensemble_mean.npy"
    rw = load_module("rw", root / "code/randomwalk_systematic_ablation.py")
    stack = load_module("stack", root / "code/stack_rank_calibration.py")
    gen = load_module("gen", root / "code/generate_post95_submission.py")
    rich = load_module("rich", root / "code/content_rich_ablation.py")
    ens = load_module("ens", root / "code/generate_randomwalk_ensemble_submission.py")

    out = root / "validation_runs" / f"dynamic_seed{args.split_seed}" / "rich_randomwalk_stack"
    sub_dir = out / "submissions"
    out.mkdir(parents=True, exist_ok=True)
    sub_dir.mkdir(parents=True, exist_ok=True)

    versions = [
        "dw_base_d128_l40_w10_win10",
        "dw_long_d128_l80_w10_win10",
        "dw_highdim_d256_l40_w10_win10",
        "dw_d256_l80_w10_win10",
        "dw_seed3407_d128_l40_w10_win10",
        "dw_graph_ap_pp",
        "n2v_p2_q1_d128_l40_w10_win10",
    ]
    cfgs = {c.version_name: c for c in rw.small_configs() + rw.graph_configs() + rw.extra_configs()}
    sys_dir = root / "validation_runs" / f"dynamic_seed{args.split_seed}" / "randomwalk_systematic"
    train_refs, pairs, y, X_base = rw.build_base_features(root, args.split_seed, main_val)
    builder = stack.ExplicitGraphFeatures(root, train_refs)
    X_rich = rich.content_rich_features(root, pairs, builder)
    blocks = []
    for version in versions:
        cfg = cfgs[version]
        model = Word2Vec.load(str(sys_dir / "models" / f"{version}.model"))
        block, _ = rw.pair_feature_block(model, pairs, cfg, root, args.split_seed, train_refs)
        blocks.append(block)
    X = np.column_stack([X_base, X_rich, *blocks, ens.aggregate(blocks)]).astype(np.float32)
    print("X", X.shape)

    rows = []
    for name, leaves in [("rich_rw7_lgb31", 31), ("rich_rw7_lgb15", 15), ("rich_rw7_lgb63", 63)]:
        oof = fit_lgb_oof(X, y, args.seed + leaves, args.n_splits, leaves=leaves)
        np.save(out / f"{name}_oof.npy", oof)
        f1, th, auc, p, r = best_f1(y, oof)
        rows.append({"stage": name, "validation_f1": f1, "threshold": th, "auc": auc, "precision": p, "recall": r, "n_features": X.shape[1]})
        print(rows[-1])
    pd.DataFrame(rows).sort_values("validation_f1", ascending=False).to_csv(out / "validation_summary.csv", index=False)

    # Generate test submissions for the best validation model.
    best = max(rows, key=lambda r: r["validation_f1"])
    best_leaves = int(best["stage"].split("lgb")[-1])
    test_pairs = np.array(gen.read_txt(root / "data_and_docs/bipartite_test_ann.txt"), dtype=np.int64)
    main_test = np.load(root / "validation_runs/dynamic_seed202/post95_test_scores/dyn202_l2d512_bpr_bigbatch_more/scores/test_vanilla_ensemble_mean.npy").astype(np.float32)
    full_refs = pd.DataFrame(gen.read_txt(root / "data_and_docs/bipartite_train_ann.txt"), columns=["source", "target"])
    test_builder = stack.ExplicitGraphFeatures(root, full_refs)
    Xht = test_builder.transform(test_pairs)
    post = load_module("post", root / "code/post95_ablation.py")
    extra = load_module("extra", root / "code/extra_score_sources_ablation.py")
    Xt = np.column_stack(
        [
            stack.add_rank_features(test_pairs, main_test),
            Xht,
            post.negative_evidence_features(Xht, main_test),
            gen.topk_content_similarity_fast(root, test_pairs, test_builder),
        ]
    ).astype(np.float32)
    selected = [Path(x.strip()) for x in (root / "validation_runs/dynamic_seed202/post95_submission/selected_variant_val_scores.txt").read_text().splitlines() if x.strip()]
    test_scores = []
    for p in selected:
        rel = p.resolve().relative_to(root / "validation_runs/dynamic_seed202")
        tp = root / "validation_runs/dynamic_seed202/post95_test_scores" / rel.parent / rel.name.replace("val_", "test_", 1)
        test_scores.append(np.load(tp).astype(np.float32))
    Xt = np.column_stack([Xt, gen.variant_feature_matrix(post, test_scores)]).astype(np.float32)
    content_test = extra.content_mean_score(root, test_pairs, test_builder)
    mf_test = np.load(root / "validation_runs/dynamic_seed202/extra_bprmf_submission/test_mf_bpr_dynamic_s202_d256_e220.npy").astype(np.float32)
    Xct, _ = extra.score_to_features(content_test, "content_mean_cos", test_pairs)
    Xmt, _ = extra.score_to_features(mf_test, "mf_bpr", test_pairs)
    Xt = np.column_stack([Xt, Xct, Xmt, rich.content_rich_features(root, test_pairs, test_builder)]).astype(np.float32)
    test_blocks = []
    for version in versions:
        cfg = cfgs[version]
        model = Word2Vec.load(str(sys_dir / "models" / f"{version}.model"))
        block, _ = rw.pair_feature_block(model, test_pairs, cfg, root, args.split_seed, full_refs)
        test_blocks.append(block)
    Xt = np.column_stack([Xt, *test_blocks, ens.aggregate(test_blocks)]).astype(np.float32)
    print("Xt", Xt.shape)
    test_score = fit_full_predict(X, y, Xt, args.seed + 500, leaves=best_leaves)
    np.save(out / f"{best['stage']}_test_pred.npy", test_score)
    known = np.load(root / "cached_scores/test_known_mask.npy").astype(bool)
    anchor = pd.read_csv(root / "validation_runs/dynamic_seed202/node2vec_deepwalk_submission/submission_content_mf_deepwalk_node2vec_lgb_th0.480000.csv")["Predicted"].to_numpy(np.int8)
    sub_rows = []
    for ratio in [0.498, 0.499, 0.500, 0.501, 0.502, float((np.load(out / f"{best['stage']}_oof.npy") >= best["threshold"]).mean())]:
        path = sub_dir / f"submission_{best['stage']}_r{ratio:.6f}.csv"
        pos, changed = write_sub(path, test_score, known, anchor, ratio)
        sub_rows.append({"path": str(path), "ratio": ratio, "test_positive_ratio": pos, "changed_vs_anchor": changed, **best})
    pd.DataFrame(sub_rows).to_csv(out / "submission_summary.csv", index=False)
    print(pd.DataFrame(sub_rows).to_string(index=False))


if __name__ == "__main__":
    main()