"""Ablate extra non-LightGCN score sources for the post95 stacker.""" from __future__ import annotations import argparse import importlib.util import pickle as pkl from pathlib import Path import lightgbm as lgb import numpy as np import pandas as pd import torch import torch.nn as nn import torch.nn.functional as F from sklearn.metrics import precision_recall_curve, roc_auc_score from sklearn.model_selection import GroupKFold, 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 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 rank01(x: np.ndarray) -> np.ndarray: order = np.argsort(x, kind="mergesort") out = np.empty(len(x), dtype=np.float32) out[order] = np.linspace(0.0, 1.0, len(x), dtype=np.float32) return out def zscore(x: np.ndarray) -> np.ndarray: return ((x - x.mean()) / (x.std() + 1e-8)).astype(np.float32) def score_to_features(scores: np.ndarray, prefix: str, pairs: np.ndarray) -> tuple[np.ndarray, list[str]]: author_rank = np.zeros(len(scores), dtype=np.float32) df = pd.DataFrame({"idx": np.arange(len(scores)), "author": pairs[:, 0], "score": scores}) for _, g in df.groupby("author", sort=False): idx = g["idx"].to_numpy() order = np.argsort(g["score"].to_numpy(), kind="mergesort") vals = np.linspace(0, 1, len(idx), dtype=np.float32) if len(idx) > 1 else np.array([1.0], dtype=np.float32) author_rank[idx[order]] = vals X = np.column_stack([scores.astype(np.float32), zscore(scores), rank01(scores), author_rank]).astype(np.float32) return X, [prefix, f"{prefix}_z", f"{prefix}_rank", f"{prefix}_author_rank"] def fit_lgb_oof(X: np.ndarray, y: np.ndarray, seed: int, n_splits: int) -> 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=31, subsample=0.9, colsample_bytree=0.9, reg_lambda=5.0, min_child_samples=80, objective="binary", verbose=-1, random_state=seed + fold, ) clf.fit(X[tr], y[tr]) oof[va] = clf.predict_proba(X[va])[:, 1] return oof def content_mean_score(root: Path, pairs: np.ndarray, builder) -> np.ndarray: cache = root / "validation_runs" / "feature_cache" cache.mkdir(parents=True, exist_ok=True) path = cache / f"content_mean_{len(pairs)}_{int(pairs[:,0].sum())}_{int(pairs[:,1].sum())}.npy" if path.exists(): return np.load(path) with (root / "data_and_docs" / "feature.pkl").open("rb") as f: feat = pkl.load(f).numpy().astype(np.float32) feat /= np.linalg.norm(feat, axis=1, keepdims=True) + 1e-8 author_mean = np.zeros((builder.num_authors, feat.shape[1]), dtype=np.float32) for a in range(builder.num_authors): hist = list(builder.author_papers[a]) if hist: v = feat[np.asarray(hist, dtype=np.int64)].mean(axis=0) author_mean[a] = v / (np.linalg.norm(v) + 1e-8) out = np.sum(author_mean[pairs[:, 0]] * feat[pairs[:, 1]], axis=1).astype(np.float32) np.save(path, out) return out class MF(nn.Module): def __init__(self, n_author: int, n_paper: int, dim: int): super().__init__() self.a = nn.Embedding(n_author, dim) self.p = nn.Embedding(n_paper, dim) self.ab = nn.Embedding(n_author, 1) self.pb = nn.Embedding(n_paper, 1) nn.init.normal_(self.a.weight, std=0.05) nn.init.normal_(self.p.weight, std=0.05) nn.init.zeros_(self.ab.weight) nn.init.zeros_(self.pb.weight) def score(self, pairs): return (self.a(pairs[:, 0]) * self.p(pairs[:, 1])).sum(-1) + self.ab(pairs[:, 0]).squeeze(-1) + self.pb(pairs[:, 1]).squeeze(-1) def train_mf_bpr_score(root: Path, train_refs: pd.DataFrame, val_pairs: pd.DataFrame, out_dir: Path, device: str, seed: int, dim: int = 256, epochs: int = 220) -> np.ndarray: out_path = out_dir / f"val_mf_bpr_s{seed}_d{dim}.npy" if out_path.exists(): return np.load(out_path) torch.manual_seed(seed) np.random.seed(seed) rng = np.random.default_rng(seed) train = train_refs[["source", "target"]].to_numpy(np.int64) val = val_pairs[["source", "target"]].to_numpy(np.int64) y = val_pairs["label"].to_numpy(np.int8) train_set = set(map(tuple, train.tolist())) model = MF(6611, 79937, dim).to(torch.device(device)) opt = torch.optim.AdamW(model.parameters(), lr=0.01, weight_decay=1e-6) train_t = torch.as_tensor(train, dtype=torch.long, device=device) val_t = torch.as_tensor(val, dtype=torch.long, device=device) batch_size = 65536 best = (-1.0, None) for ep in range(epochs): idx = torch.randint(0, train_t.size(0), (batch_size,), device=device) pos = train_t[idx] neg_np = np.empty((batch_size, 2), dtype=np.int64) authors = pos[:, 0].detach().cpu().numpy() filled = 0 while filled < batch_size: papers = rng.integers(0, 79937, size=batch_size - filled) for a, p in zip(authors[filled:], papers): if (int(a), int(p)) not in train_set: neg_np[filled] = (a, p) filled += 1 if filled >= batch_size: break neg = torch.as_tensor(neg_np, dtype=torch.long, device=device) loss = -F.logsigmoid(model.score(pos) - model.score(neg)).mean() opt.zero_grad() loss.backward() opt.step() if (ep + 1) % 20 == 0 or ep == epochs - 1: with torch.no_grad(): scores = [] for st in range(0, len(val), 131072): scores.append(model.score(val_t[st : st + 131072]).detach().cpu().numpy()) scores = np.concatenate(scores).astype(np.float32) f1, th, auc, _, _ = best_f1(y, scores) if f1 > best[0]: best = (f1, scores.copy()) print(f"mf epoch={ep+1:03d} loss={loss.item():.4f} f1={f1:.6f} th={th:.6f} auc={auc:.6f}") np.save(out_path, best[1]) return best[1] def train_ranker_oof(X: np.ndarray, y: np.ndarray, pairs: np.ndarray, seed: int, out_dir: Path) -> np.ndarray: out_path = out_dir / "val_lgbmranker_oof.npy" if out_path.exists(): return np.load(out_path) oof = np.zeros(len(y), dtype=np.float32) gkf = GroupKFold(n_splits=5) groups = pairs[:, 0] for fold, (tr, va) in enumerate(gkf.split(X, y, groups=groups), start=1): tr_order = np.lexsort((np.arange(len(tr)), pairs[tr, 0])) va_order = np.lexsort((np.arange(len(va)), pairs[va, 0])) tr_idx = tr[tr_order] va_idx = va[va_order] tr_group = pd.Series(pairs[tr_idx, 0]).value_counts(sort=False).to_numpy() ranker = lgb.LGBMRanker( objective="lambdarank", metric="ndcg", n_estimators=700, learning_rate=0.03, num_leaves=31, subsample=0.9, colsample_bytree=0.9, reg_lambda=10.0, min_child_samples=60, random_state=seed + fold, verbose=-1, ) ranker.fit(X[tr_idx], y[tr_idx], group=tr_group) oof[va_idx] = ranker.predict(X[va_idx]).astype(np.float32) print(f"ranker fold={fold} f1={best_f1(y[va_idx], oof[va_idx])[0]:.6f}") np.save(out_path, oof) return oof def main() -> None: parser = argparse.ArgumentParser() parser.add_argument("--package-root", type=Path, default=Path(__file__).resolve().parents[1]) parser.add_argument("--split-seed", type=int, default=202) parser.add_argument("--main-val-score-file", type=Path, required=True) parser.add_argument("--device", default="cuda:0" if torch.cuda.is_available() else "cpu") parser.add_argument("--seed", type=int, default=202) parser.add_argument("--n-splits", type=int, default=5) parser.add_argument("--skip-mf", action="store_true") args = parser.parse_args() root = args.package_root stack = load_module("stack", root / "code" / "stack_rank_calibration.py") lgcn = load_module("lgcn", root / "code" / "train_val_lgcn_ensemble.py") post = load_module("post", root / "code" / "post95_ablation.py") gen = load_module("gen", root / "code" / "generate_post95_submission.py") train_refs, val_pairs = lgcn.make_notebook_style_split(root, args.split_seed, 0.9) pairs = val_pairs[["source", "target"]].to_numpy(np.int64) y = val_pairs["label"].to_numpy(np.int8) main = np.load(args.main_val_score_file).astype(np.float32) out_dir = root / "validation_runs" / f"dynamic_seed{args.split_seed}" / "extra_score_sources" out_dir.mkdir(parents=True, exist_ok=True) print("building post95 base features") builder = stack.ExplicitGraphFeatures(root, train_refs) X_hand = builder.transform(pairs) X = np.column_stack( [ stack.add_rank_features(pairs, main), X_hand, post.negative_evidence_features(X_hand, main), gen.topk_content_similarity_fast(root, pairs, builder), ] ).astype(np.float32) selected = [Path(x.strip()) for x in (root / "validation_runs" / f"dynamic_seed{args.split_seed}" / "post95_submission" / "selected_variant_val_scores.txt").read_text().splitlines() if x.strip()] X = np.column_stack([X, gen.variant_feature_matrix(post, [np.load(p).astype(np.float32) for p in selected])]).astype(np.float32) rows = [] base_oof = fit_lgb_oof(X, y, args.seed, args.n_splits) f1, th, auc, p, r = best_f1(y, base_oof) rows.append({"stage": "post95_lgbm_baseline", "f1": f1, "threshold": th, "auc": auc, "precision": p, "recall": r, "n_features": X.shape[1]}) np.save(out_dir / "post95_lgbm_baseline_oof.npy", base_oof) extra_blocks = [] extra_names = [] print("adding pure content mean-cos score") content = content_mean_score(root, pairs, builder) Xc, names = score_to_features(content, "content_mean_cos", pairs) extra_blocks.append(Xc) extra_names.extend(names) X_cur = np.column_stack([X, *extra_blocks]).astype(np.float32) oof = fit_lgb_oof(X_cur, y, args.seed + 10, args.n_splits) f1, th, auc, p, r = best_f1(y, oof) rows.append({"stage": "+content_mean_cos", "f1": f1, "threshold": th, "auc": auc, "precision": p, "recall": r, "n_features": X_cur.shape[1]}) np.save(out_dir / "content_mean_cos_stack_oof.npy", oof) if not args.skip_mf: print("training/adding BPR-MF score") mf = train_mf_bpr_score(root, train_refs, val_pairs, out_dir, args.device, args.seed) Xm, names = score_to_features(mf, "mf_bpr", pairs) extra_blocks.append(Xm) extra_names.extend(names) X_cur = np.column_stack([X, *extra_blocks]).astype(np.float32) oof = fit_lgb_oof(X_cur, y, args.seed + 20, args.n_splits) f1, th, auc, p, r = best_f1(y, oof) rows.append({"stage": "+bpr_mf", "f1": f1, "threshold": th, "auc": auc, "precision": p, "recall": r, "n_features": X_cur.shape[1]}) np.save(out_dir / "bpr_mf_stack_oof.npy", oof) print("training/adding author-group LGBMRanker OOF score") ranker_scores = train_ranker_oof(X, y, pairs, args.seed, out_dir) Xr, names = score_to_features(ranker_scores, "lgbmranker_author_oof", pairs) extra_blocks.append(Xr) extra_names.extend(names) X_cur = np.column_stack([X, *extra_blocks]).astype(np.float32) oof = fit_lgb_oof(X_cur, y, args.seed + 30, args.n_splits) f1, th, auc, p, r = best_f1(y, oof) rows.append({"stage": "+lgbmranker_author", "f1": f1, "threshold": th, "auc": auc, "precision": p, "recall": r, "n_features": X_cur.shape[1]}) np.save(out_dir / "lgbmranker_stack_oof.npy", oof) pd.Series(extra_names).to_csv(out_dir / "extra_feature_names.csv", index=False) result = pd.DataFrame(rows).sort_values("f1", ascending=False) result.to_csv(out_dir / "extra_score_ablation.csv", index=False) print(result.to_string(index=False)) if __name__ == "__main__": main()