reference_data_preprocess.py

import os
import json
import gzip
import pickle
import logging
from collections import defaultdict
from typing import Dict, List, Tuple, Optional
import pandas as pd
from tqdm import tqdm
import numpy as np
from scipy.sparse import lil_matrix, save_npz
from huggingface_hub import HfApi, hf_hub_download

def _open_any(path: str):
    if path.endswith(".gz"):
        return gzip.open(path, "rt", encoding="utf-8")
    return open(path, "r", encoding="utf-8")

def _clean_title(s: str) -> str:
    if not isinstance(s, str):
        return ""
    import html as _html
    s = _html.unescape(s)
    return " ".join(s.split())

def _load_asin2title(meta_file_path: str) -> Dict[str, str]:
    """meta_{dataset}.json(.gz) line-delimited JSON에서 asin→cleaned title 맵 생성"""
    asin2title = {}
    with _open_any(meta_file_path) as f:
        for line in f:
            line = line.strip()
            if not line:
                continue
            try:
                d = json.loads(line)
            except json.JSONDecodeError:
                continue
            # parent_asin 또는 asin 키 모두 지원
            asin = d.get("parent_asin") or d.get("asin")
            title = _clean_title(d.get("title", ""))
            if asin:
                asin2title[asin] = (title if title else "empty title")
    return asin2title

# ---------------------------------------------------

def _download_hf_csv(repo_id: str, path: str, repo_type: str = "dataset") -> str:
    return hf_hub_download(repo_id=repo_id, repo_type=repo_type, filename=path)


def _load_hf_last_out_splits(
    fname: str,
    core: str,
    repo_id: str,
    user_col: str,
    item_col: str,
    time_col: str,
    rating_col: Optional[str] = None,
    logger=None,
):
    log = logger or logging.getLogger(__name__)
    api = HfApi()
    log.info("Listing HF repo files for last_out splits...")
    files = set(api.list_repo_files(repo_id=repo_id, repo_type="dataset"))

    prefix = f"benchmark/{core}/last_out/"
    split_paths = {
        "train": f"{prefix}{fname}.train.csv",
        "valid": f"{prefix}{fname}.valid.csv",
        "test": f"{prefix}{fname}.test.csv",
    }

    dfs = {}
    for split, path in split_paths.items():
        if path not in files:
            raise RuntimeError(f"Missing HF file: {path}")
        log.info(f"Downloading split file: {path}")
        local_path = _download_hf_csv(repo_id, path, repo_type="dataset")
        df = pd.read_csv(local_path)
        log.info(f"Loaded split {split}: {len(df)} rows")

        keep_cols = [user_col, item_col, time_col]
        if rating_col and rating_col in df.columns:
            keep_cols.append(rating_col)

        missing = [c for c in keep_cols if c not in df.columns]
        if missing:
            raise ValueError(f"Missing columns {missing} in {path}. cols={list(df.columns)}")

        df = df[keep_cols].copy()
        rename_map = {
            user_col: "user_id",
            item_col: "parent_asin",
            time_col: "timestamp",
        }
        if rating_col:
            rename_map[rating_col] = "rating"
        df = df.rename(columns=rename_map)
        dfs[split] = df

    return dfs


def _build_rating_lookup_from_reviews(
    review_json_path: str,
    keys: set[tuple[str, str, int]],
    logger=None,
):
    log = logger or logging.getLogger(__name__)
    rating_map: dict[tuple[str, str, int], int] = {}

    if not keys:
        return rating_map

    with _open_any(review_json_path) as f:
        for line in tqdm(f, desc="Building rating lookup from reviews"):
            line = line.strip()
            if not line:
                continue
            try:
                d = json.loads(line)
            except json.JSONDecodeError:
                continue

            uid = d.get("reviewerID")
            asin = d.get("asin")
            ts = d.get("unixReviewTime")
            if uid is None or asin is None or ts is None:
                continue

            key = (str(uid), str(asin), int(ts))
            if key not in keys:
                continue

            score = d.get("overall", d.get("rating", d.get("stars", 0)))
            try:
                rating = int(float(score))
            except Exception:
                rating = 0
            rating_map[key] = rating

    log.info(f"Loaded ratings for {len(rating_map)} interactions.")
    return rating_map


def _build_user_sequences(df: pd.DataFrame) -> dict[str, list[tuple[str, int, int]]]:
    seqs: dict[str, list[tuple[str, int, int]]] = {}
    for uid, sub in df.groupby("user_id", sort=False):
        sub = sub.sort_values("timestamp")
        items = sub["parent_asin"].astype(str).tolist()
        times = sub["timestamp"].astype(int).tolist()
        ratings = sub["rating"].astype(int).tolist()
        seqs[str(uid)] = list(zip(items, times, ratings))
    return seqs


def process_hf_last_out_to_llara(
    fname: str,
    dataset: str,
    output_dir: str,
    core: str = "5core",
    repo_id: str = "McAuley-Lab/Amazon-Reviews-2023",
    user_col: str = "user_id",
    item_col: str = "parent_asin",
    time_col: str = "timestamp",
    rating_col: Optional[str] = None,
    review_json_path: Optional[str] = None,
    meta_file_path: Optional[str] = None,
    default_rating: int = 0,
    logger=None,
):
    """
    HF last_out(train/valid/test)을 그대로 사용해 LLaRA 입력(train_one/val/test + item_meta) 생성.
    """
    log = logger or logging.getLogger(__name__)
    os.makedirs(os.path.join(output_dir, dataset), exist_ok=True)
    dataset_output_dir = os.path.join(output_dir, dataset)

    # meta 파일 경로: 로컬에 없으면 HuggingFace에서 다운로드
    if meta_file_path is None:
        meta_file_path = os.path.join(os.path.dirname(output_dir), f"meta_{dataset}.json")
        if not os.path.exists(meta_file_path) and os.path.exists(meta_file_path + ".gz"):
            meta_file_path = meta_file_path + ".gz"

    if meta_file_path is None or not os.path.exists(meta_file_path):
        # HuggingFace에서 meta 파일 다운로드 시도
        hf_meta_filename = f"raw/meta_categories/meta_{dataset}.jsonl"
        log.info(f"Downloading meta file from HuggingFace: {repo_id}/{hf_meta_filename}")
        try:
            meta_file_path = hf_hub_download(
                repo_id=repo_id,
                repo_type="dataset",
                filename=hf_meta_filename,
            )
            log.info(f"Meta file downloaded to: {meta_file_path}")
        except Exception as e:
            log.warning(f"Failed to download meta file from HF: {e}  (title은 'empty title'로 대체)")
            meta_file_path = None

    if meta_file_path is None or not os.path.exists(meta_file_path):
        log.warning(f"Meta file not found (title은 'empty title'로 대체)")
        asin2title = {}
    else:
        asin2title = _load_asin2title(meta_file_path)

    splits = _load_hf_last_out_splits(
        fname=fname,
        core=core,
        repo_id=repo_id,
        user_col=user_col,
        item_col=item_col,
        time_col=time_col,
        rating_col=rating_col,
        logger=log,
    )

    for split, df in splits.items():
        df = df.dropna(subset=["user_id", "parent_asin", "timestamp"]).copy()
        df["timestamp"] = pd.to_numeric(df["timestamp"], errors="coerce")
        df = df.dropna(subset=["timestamp"]).copy()
        df["timestamp"] = df["timestamp"].astype(np.int64)
        splits[split] = df

    # review data로 rating 채우기 (필요 시)
    if review_json_path:
        key_set = set()
        for df in splits.values():
            for uid, asin, ts in zip(df["user_id"], df["parent_asin"], df["timestamp"]):
                key_set.add((str(uid), str(asin), int(ts)))

        rating_map = _build_rating_lookup_from_reviews(review_json_path, key_set, logger=log)
        for df in splits.values():
            df["rating"] = [
                int(rating_map.get((str(u), str(a), int(t)), default_rating))
                for u, a, t in zip(df["user_id"], df["parent_asin"], df["timestamp"])
            ]
    else:
        if rating_col and "rating" in splits["train"].columns:
            for df in splits.values():
                df["rating"] = pd.to_numeric(df["rating"], errors="coerce").fillna(default_rating).astype(int)
        else:
            for df in splits.values():
                df["rating"] = int(default_rating)

    # user/item id mapping (string -> int)
    all_users = sorted(set(pd.concat([splits[s]["user_id"] for s in splits]).astype(str)))
    all_items = sorted(set(pd.concat([splits[s]["parent_asin"] for s in splits]).astype(str)))

    user_id_to_idx = {u: i for i, u in enumerate(all_users)}
    asin_to_idx = {a: i for i, a in enumerate(all_items)}

    # item_meta.parquet (id, title)
    meta_rows = []
    for asin, idx in asin_to_idx.items():
        title = asin2title.get(asin, "empty title") if asin2title else "empty title"
        meta_rows.append({"id": idx, "title": title})
    meta_df = pd.DataFrame(meta_rows)
    meta_df.to_parquet(os.path.join(dataset_output_dir, "item_meta.parquet"), index=False)

    # build per-user sequences
    train_seqs = _build_user_sequences(splits["train"])
    val_seqs = _build_user_sequences(splits["valid"])
    test_seqs = _build_user_sequences(splits["test"])

    train_rows = []
    val_rows = []
    test_rows = []

    for u in all_users:
        uid = str(u)
        uidx = user_id_to_idx[uid]

        train_items = [asin_to_idx[a] for a, _, _ in train_seqs.get(uid, [])]
        train_ratings = [r for _, _, r in train_seqs.get(uid, [])]

        val_items = [asin_to_idx[a] for a, _, _ in val_seqs.get(uid, [])]
        val_ratings = [r for _, _, r in val_seqs.get(uid, [])]

        test_items = [asin_to_idx[a] for a, _, _ in test_seqs.get(uid, [])]
        test_ratings = [r for _, _, r in test_seqs.get(uid, [])]

        # train (one per user)
        if len(train_items) >= 2:
            seq = train_items[:-1]
            seq_rating = train_ratings[:-1] if train_ratings else [default_rating] * len(seq)
            target = train_items[-1]
            target_rating = train_ratings[-1] if train_ratings else default_rating
            train_rows.append(
                {
                    "user_id": uidx,
                    "seq": seq,
                    "next": int(target),
                    "len_seq": int(len(seq)),
                    "seq_rating": seq_rating,
                    "target_rating": int(target_rating),
                }
            )

        # val (one per user)
        if len(val_items) >= 1:
            hist = train_items + val_items[:-1]
            hist_r = train_ratings + val_ratings[:-1]
            if not hist_r:
                hist_r = [default_rating] * len(hist)
            target = val_items[-1]
            target_rating = val_ratings[-1] if val_ratings else default_rating
            val_rows.append(
                {
                    "user_id": uidx,
                    "seq": hist,
                    "next": int(target),
                    "len_seq": int(len(hist)),
                    "seq_rating": hist_r,
                    "target_rating": int(target_rating),
                }
            )

        # test (one per user)
        if len(test_items) >= 1:
            hist = train_items + val_items + test_items[:-1]
            hist_r = train_ratings + val_ratings + test_ratings[:-1]
            if not hist_r:
                hist_r = [default_rating] * len(hist)
            target = test_items[-1]
            target_rating = test_ratings[-1] if test_ratings else default_rating
            test_rows.append(
                {
                    "user_id": uidx,
                    "seq": hist,
                    "next": int(target),
                    "len_seq": int(len(hist)),
                    "seq_rating": hist_r,
                    "target_rating": int(target_rating),
                }
            )

    train_df = pd.DataFrame(train_rows)
    val_df = pd.DataFrame(val_rows)
    test_df = pd.DataFrame(test_rows)

    train_df.to_pickle(os.path.join(dataset_output_dir, "train_one_data.df"))
    val_df.to_pickle(os.path.join(dataset_output_dir, "val_data.df"))
    test_df.to_pickle(os.path.join(dataset_output_dir, "test_data.df"))

    # mapping files for reference
    with open(os.path.join(dataset_output_dir, f"{dataset}_user_id_to_idx.json"), "w") as f:
        json.dump(user_id_to_idx, f)
    with open(os.path.join(dataset_output_dir, f"{dataset}_item_id_to_idx.json"), "w") as f:
        json.dump(asin_to_idx, f)

    log.info("HF last_out preprocessing finished.")
    return True

def process_data(
    json_file_path: str,
    dataset: str,
    output_dir: str,
    count_threshold: int = 5,
    logger=None,
    meta_file_path: str = None
):
    """
    아이템 식별자는 asin.
    idx는 asin을 정렬/열거하여 부여하고, title을 매핑(없으면 'empty title').
    """
    log = logger or logging.getLogger(__name__)
    os.makedirs(os.path.join(output_dir, dataset), exist_ok=True)
    dataset_output_dir = os.path.join(output_dir, dataset)

    # meta 파일 경로 기본값 추정
    if meta_file_path is None:
        meta_file_path = os.path.join(os.path.dirname(json_file_path), f"meta_{dataset}.json")
        if not os.path.exists(meta_file_path) and os.path.exists(meta_file_path + ".gz"):
            meta_file_path = meta_file_path + ".gz"

    if not os.path.exists(meta_file_path):
        log.warning(f"Meta file not found: {meta_file_path}  (title은 'empty title'로 대체)")
        asin2title = {}
    else:
        asin2title = _load_asin2title(meta_file_path)

    # 1) Pass1: overall>=3 기준으로 사용자/아이템(asin) 카운팅
    user_cnt = defaultdict(int)
    item_cnt = defaultdict(int)
    try:
        with _open_any(json_file_path) as f:
            for line in tqdm(f, desc="Pass1: counting (overall>=3)"):
                try:
                    d = json.loads(line)
                except json.JSONDecodeError:
                    continue
                try:
                    score = d.get("overall", d.get("rating", d.get("stars", 0)))
                    score = float(score)
                except Exception:
                    score = 0
                if score < 3:
                    continue
                uid = d.get("reviewerID")
                iid_asin = d.get("asin")
                if uid is None or iid_asin is None:
                    continue
                user_cnt[uid] += 1
                item_cnt[iid_asin] += 1
    except Exception as e:
        log.error(f"Error during counting pass: {e}")
        return False

    # 2) Pass2: threshold 통과 상호작용 수집 (asin 기준)
    interactions = []
    try:
        with _open_any(json_file_path) as f:
            for line in tqdm(f, desc="Pass2: collect (threshold ok, no rating filter)"):
                try:
                    d = json.loads(line)
                except json.JSONDecodeError:
                    continue
                uid = d.get("reviewerID")
                iid_asin = d.get("asin")
                t   = d.get("unixReviewTime")
                if uid is None or iid_asin is None or t is None:
                    continue
                if user_cnt[uid] < count_threshold or item_cnt[iid_asin] < count_threshold:
                    continue
                title = _clean_title(asin2title.get(iid_asin, "")) if asin2title else ""
                if not title:
                    title = "empty title"
                interactions.append({"user": uid, "asin": iid_asin, "title": title, "time": int(t)})
    except Exception as e:
        log.error(f"Error during filtering pass: {e}")
        return False

    if not interactions:
        log.error("No interactions left after threshold filtering.")
        return False

    data_df = pd.DataFrame(interactions)
    log.info(f"Collected {len(data_df)} interactions after filtering.")

    # 3) 유저별 시간 정렬 시퀀스 (asin, time)
    interactions_by_user_timed: Dict[str, List[Tuple[str, int]]] = (
        data_df.groupby("user", sort=False)
        .apply(lambda x: sorted(list(zip(x["asin"], x["time"])), key=lambda k: k[1]))
        .to_dict()
    )

    # 4) asin 기반 인덱스 부여
    final_users = list(interactions_by_user_timed.keys())
    final_asins = {asin for seq in interactions_by_user_timed.values() for asin, _ in seq}

    user_id_list = sorted(final_users)
    asin_list    = sorted(final_asins)

    user_id_to_idx = {u: i for i, u in enumerate(user_id_list)}
    asin_to_idx    = {a: i for i, a in enumerate(asin_list)}
    idx_to_asin    = {i: a for a, i in asin_to_idx.items()}

    # idx → title 매핑 만들기 (없으면 'empty title')
    idx_to_title = {i: (asin2title.get(idx_to_asin[i], "empty title")) for i in range(len(asin_list))}

    num_users = len(user_id_list)
    num_items = len(asin_list)
    log.info(f"[ASIN-INDEXED] Final num_users={num_users}, num_items={num_items}")

    # 5) 행렬/시퀀스 구성 (열=asin index)
    interaction_matrix = lil_matrix((num_users, num_items), dtype=np.float32)
    eval_sequences: Dict[int, List[int]] = {}

    for user_id, seq_tuples in tqdm(interactions_by_user_timed.items(), desc="Build matrix & sequences (asin-indexed)"):
        uidx = user_id_to_idx[user_id]
        item_idx_sequence = []
        for asin, _t in seq_tuples:
            if asin in asin_to_idx:
                item_idx_sequence.append(asin_to_idx[asin])
        if not item_idx_sequence:
            continue
        eval_sequences[uidx] = item_idx_sequence
        for itx in item_idx_sequence[:-1]:
            interaction_matrix[uidx, itx] = 1.0

    log.info(f"Final interaction count for matrix: {interaction_matrix.nnz}")
    log.info(f"Final matrix shape: ({num_users}, {num_items})")
    print(f"Final matrix shape: ({num_users}, {num_items})")
    if not eval_sequences:
        log.error("Failed to generate any evaluation sequences (asin-indexed).")
        return False

    # 6) 아이템(asin idx) 인기 저장
    #    - 모든 인터랙션에서 등장한 아이템을 카운트(마지막 타깃 포함)
    #    - 등장하지 않은 아이템도 0으로 채워 전체 아이템 풀(0..num_items-1) 커버
    try:
        from collections import Counter
        pop_counter = Counter()

        # interactions_by_user_timed: user -> [(asin, time), ...]
        for seq in interactions_by_user_timed.values():
            for asin, _t in seq:
                if asin in asin_to_idx:
                    pop_counter[asin_to_idx[asin]] += 1

        # 전체 아이템 인덱스(0-based) 범위에 대해 0 채움
        popularity_dict = {i: int(pop_counter.get(i, 0)) for i in range(num_items)}

        popularity_path = os.path.join(dataset_output_dir, f"{dataset}_item_popularity.json")
        with open(popularity_path, "w") as f:
            json.dump(popularity_dict, f)
        log.info(f"[Beauty-style] Item popularity saved to: {popularity_path}")
    except Exception as e:
        log.warning(f"Saving item popularity failed (continuing): {e}")

    # 7) 저장 (asin-indexed)
    try:
        # 시퀀스
        seq_path = os.path.join(dataset_output_dir, f"{dataset}_eval_sequences.pkl")
        with open(seq_path, "wb") as f:
            pickle.dump(eval_sequences, f)
        log.info(f"Evaluation sequences saved to: {seq_path}")

        # 행렬
        matrix_path = os.path.join(dataset_output_dir, f"{dataset}_interaction_matrix.npz")
        save_npz(matrix_path, interaction_matrix.tocsr())
        log.info(f"Interaction matrix saved to: {matrix_path}")

        # 유저 맵
        user_map_path = os.path.join(dataset_output_dir, f"{dataset}_user_id_to_idx.json")
        with open(user_map_path, "w") as f:
            json.dump(user_id_to_idx, f)
        log.info(f"User ID map saved to: {user_map_path}")

        # 아이템 맵들 (asin/idx/title)
        item_id_map_path = os.path.join(dataset_output_dir, f"{dataset}_item_id_to_idx.json")  # asin→idx
        with open(item_id_map_path, "w") as f:
            json.dump(asin_to_idx, f)
        log.info(f"ASIN→idx map saved to: {item_id_map_path}")

        idx_to_asin_path = os.path.join(dataset_output_dir, f"{dataset}_idx_to_asin.json")
        with open(idx_to_asin_path, "w") as f:
            json.dump(idx_to_asin, f)
        log.info(f"idx→ASIN map saved to: {idx_to_asin_path}")

        idx_to_title_path = os.path.join(dataset_output_dir, f"{dataset}_idx_to_title.json")
        with open(idx_to_title_path, "w", encoding="utf-8") as f:
            json.dump({int(k): v for k, v in idx_to_title.items()}, f, ensure_ascii=False)
        log.info(f"idx→title map saved to: {idx_to_title_path}")

        # TSV: index \t asin \t title  (후속 파이프라인 호환용)
        tsv_map_path = os.path.join(dataset_output_dir, f"{dataset}_item_index_to_title.txt")
        with open(tsv_map_path, "w", encoding="utf-8") as f:
            f.write("index\tasin\ttitle\n")
            for idx in range(num_items):
                asin = idx_to_asin[idx]
                title = idx_to_title[idx]
                f.write(f"{idx}\t{asin}\t{title}\n")
        log.info(f"Item index/asin/title TSV saved to: {tsv_map_path}")

    except Exception as e:
        log.error(f"Error during file saving: {e}")
        return False

    log.info("Data processing (asin-indexed) finished successfully.")
    return True

if __name__ == "__main__":
    # Example: HF last_out -> LLaRA input
    dataset = "Industrial_and_Scientific"
    output_dir = "/mnt/data0/woosung/KDD2026/LLaRA/data"

    print(f"Start HF last_out processing {dataset} ...")
    success = process_hf_last_out_to_llara(
        fname=dataset,
        dataset=dataset,
        output_dir=output_dir,
        core="5core",
        repo_id="McAuley-Lab/Amazon-Reviews-2023",
        user_col="user_id",
        item_col="parent_asin",
        time_col="timestamp",
        rating_col=None,
        review_json_path=None,
        meta_file_path=None,
        default_rating=0,
    )
    if success:
        print("HF last_out processing finished!")
    else:
        print("HF last_out processing failed!")