File size: 15,193 Bytes
1359487 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 | """
Feature engineering and dataset preparation for the recommendation pipeline.
Steps:
1. Encode user/movie IDs to contiguous 0-based integers
2. Compute user features (genre preferences, activity stats)
3. Compute item features (genre encoding, genome PCA embeddings, year)
4. Temporal train/val/test split
5. Generate negative samples for training
"""
import logging
import pickle
from pathlib import Path
from typing import Optional
import numpy as np
import pandas as pd
from sklearn.decomposition import TruncatedSVD
from sklearn.preprocessing import StandardScaler
from .loader import ALL_GENRES, load_all
logger = logging.getLogger(__name__)
GENRE_COLS = [
f"g_{g.replace('-','_').replace('(','').replace(')','').replace(' ','_')}"
for g in ALL_GENRES
]
class Preprocessor:
"""
Encapsulates all feature-engineering logic.
Call fit() on training data, then transform() on any split.
Artifacts are saved to output_dir for serving-time use.
"""
def __init__(self, genome_n_components: int = 32):
self.genome_n_components = genome_n_components
self.user_id_map: dict[int, int] = {}
self.movie_id_map: dict[int, int] = {}
self.user_id_unmap: dict[int, int] = {}
self.movie_id_unmap: dict[int, int] = {}
self.num_users: int = 0
self.num_movies: int = 0
self.genre_scaler = StandardScaler()
self.genome_svd: Optional[TruncatedSVD] = None
self.item_features: Optional[np.ndarray] = None # (num_movies, feature_dim)
self.user_features: Optional[np.ndarray] = None # (num_users, feature_dim)
self.movie_df: Optional[pd.DataFrame] = None
self.fitted = False
# ------------------------------------------------------------------
# ID encoding
# ------------------------------------------------------------------
def _build_id_maps(self, ratings: pd.DataFrame, movies: pd.DataFrame) -> None:
unique_users = sorted(ratings["userId"].unique())
unique_movies = sorted(movies["movieId"].unique())
self.user_id_map = {uid: idx for idx, uid in enumerate(unique_users)}
self.movie_id_map = {mid: idx for idx, mid in enumerate(unique_movies)}
self.user_id_unmap = {v: k for k, v in self.user_id_map.items()}
self.movie_id_unmap = {v: k for k, v in self.movie_id_map.items()}
self.num_users = len(unique_users)
self.num_movies = len(unique_movies)
logger.info(f"ID maps: {self.num_users:,} users, {self.num_movies:,} movies")
def encode_ratings(self, ratings: pd.DataFrame) -> pd.DataFrame:
df = ratings.copy()
df["user_idx"] = df["userId"].map(self.user_id_map)
df["movie_idx"] = df["movieId"].map(self.movie_id_map)
# Drop rows where IDs aren't in training vocabulary
df = df.dropna(subset=["user_idx", "movie_idx"])
df["user_idx"] = df["user_idx"].astype(np.int32)
df["movie_idx"] = df["movie_idx"].astype(np.int32)
return df
# ------------------------------------------------------------------
# Item feature matrix
# ------------------------------------------------------------------
def _build_item_features(
self, movies: pd.DataFrame, genome_scores: Optional[pd.DataFrame]
) -> np.ndarray:
"""Returns (num_movies, feature_dim) float32 array."""
# Genre multi-hot β shape (num_movies, 20)
genre_matrix = np.zeros((self.num_movies, len(GENRE_COLS)), dtype=np.float32)
for _, row in movies.iterrows():
idx = self.movie_id_map.get(row["movieId"])
if idx is None:
continue
for i, col in enumerate(GENRE_COLS):
genre_matrix[idx, i] = row.get(col, 0)
# Normalised release year β shape (num_movies, 1)
year_vec = np.zeros((self.num_movies, 1), dtype=np.float32)
for _, row in movies.iterrows():
idx = self.movie_id_map.get(row["movieId"])
if idx is None:
continue
year_vec[idx, 0] = max(0, row.get("year", 2000) - 1900) / 100.0
features = [genre_matrix, year_vec]
# Genome tag embeddings via SVD β shape (num_movies, genome_n_components)
if genome_scores is not None:
genome_matrix = self._build_genome_matrix(movies, genome_scores)
features.append(genome_matrix)
item_feats = np.concatenate(features, axis=1)
logger.info(f"Item feature dim: {item_feats.shape[1]}")
return item_feats
def _build_genome_matrix(
self, movies: pd.DataFrame, genome_scores: pd.DataFrame
) -> np.ndarray:
"""Pivot genome scores to (num_movies, num_tags), then reduce with SVD."""
pivot = genome_scores.pivot_table(
index="movieId", columns="tagId", values="relevance", fill_value=0.0
)
# Only keep movies in our vocabulary
valid_ids = [
mid for mid in pivot.index if mid in self.movie_id_map
]
pivot = pivot.loc[valid_ids]
if self.genome_svd is None:
n = min(self.genome_n_components, pivot.shape[1] - 1, pivot.shape[0] - 1)
self.genome_svd = TruncatedSVD(n_components=n, random_state=42)
reduced = self.genome_svd.fit_transform(pivot.values)
else:
reduced = self.genome_svd.transform(pivot.values)
genome_matrix = np.zeros(
(self.num_movies, self.genome_svd.n_components), dtype=np.float32
)
for orig_id, row in zip(valid_ids, reduced):
idx = self.movie_id_map[orig_id]
genome_matrix[idx] = row.astype(np.float32)
logger.info(f"Genome SVD shape: {genome_matrix.shape}")
return genome_matrix
# ------------------------------------------------------------------
# User feature matrix
# ------------------------------------------------------------------
def _build_user_features(
self, ratings: pd.DataFrame, encoded: pd.DataFrame
) -> np.ndarray:
"""
Returns (num_users, feature_dim) float32 array.
Features: genre preferences (20), avg_rating (1), log_count (1) β 22-dim.
Fully vectorised: uses np.bincount (compiled C) per genre column instead
of a Python groupby loop, reducing runtime from minutes to seconds on 2M rows.
"""
n_genres = len(GENRE_COLS)
genre_pref = np.zeros((self.num_users, n_genres), dtype=np.float64)
avg_rating = np.zeros((self.num_users, 1), dtype=np.float32)
log_count = np.zeros((self.num_users, 1), dtype=np.float32)
if self.item_features is not None:
user_idxs = encoded["user_idx"].values.astype(np.int64)
movie_idxs = encoded["movie_idx"].values.astype(np.int64)
rating_vals = encoded["rating"].values.astype(np.float64)
item_genre = self.item_features[:, :n_genres].astype(np.float64)
# Weighted genre signal per interaction: [N, n_genres]
genre_weighted = item_genre[movie_idxs] * rating_vals[:, None]
# Accumulate per user with np.bincount (fast C loop, one call per genre)
for g in range(n_genres):
genre_pref[:, g] = np.bincount(
user_idxs, weights=genre_weighted[:, g], minlength=self.num_users
)
# Per-user interaction counts and rating sums
counts = np.bincount(user_idxs, minlength=self.num_users).astype(np.float64)
rating_sum = np.bincount(user_idxs, weights=rating_vals, minlength=self.num_users)
mask = counts > 0
genre_pref[mask] /= counts[mask, None]
avg_rating[mask, 0] = (rating_sum[mask] / counts[mask] / 5.0).astype(np.float32)
log_count[mask, 0] = (np.log1p(counts[mask]) / 10.0).astype(np.float32)
user_feats = np.concatenate([genre_pref.astype(np.float32), avg_rating, log_count], axis=1)
logger.info(f"User feature dim: {user_feats.shape[1]}")
return user_feats
# ------------------------------------------------------------------
# Temporal split
# ------------------------------------------------------------------
def temporal_split(
self,
encoded: pd.DataFrame,
val_frac: float = 0.10,
test_frac: float = 0.10,
) -> tuple[pd.DataFrame, pd.DataFrame, pd.DataFrame]:
"""
Split by timestamp: oldest interactions β train, newest β test.
This avoids data leakage from the future.
"""
n = len(encoded)
test_start = int(n * (1 - test_frac))
val_start = int(n * (1 - val_frac - test_frac))
train = encoded.iloc[:val_start].copy()
val = encoded.iloc[val_start:test_start].copy()
test = encoded.iloc[test_start:].copy()
logger.info(
f"Split β train: {len(train):,}, val: {len(val):,}, test: {len(test):,}"
)
return train, val, test
# ------------------------------------------------------------------
# Negative sampling
# ------------------------------------------------------------------
def generate_negatives(
self,
encoded: pd.DataFrame,
n_neg: int = 4,
seed: int = 42,
) -> pd.DataFrame:
"""
For each positive (user, movie) pair, sample n_neg random negatives.
Returns DataFrame with columns: user_idx, pos_movie_idx, neg_movie_idx.
Fully vectorised: samples negatives in one numpy call, then fixes the rare
case where a sampled negative equals the positive (collision rate β 0.5% on
MovieLens given ~145 ratings per user vs 27K movies). Full seen-set filtering
is skipped intentionally β the sparsity of the dataset means almost all random
items are true negatives, which is standard practice for large-scale BPR.
"""
rng = np.random.default_rng(seed)
user_idxs = encoded["user_idx"].values.astype(np.int32)
pos_idxs = encoded["movie_idx"].values.astype(np.int32)
N = len(user_idxs)
# Sample n_neg negatives per interaction in one vectorised call
neg_samples = rng.integers(0, self.num_movies, size=(N, n_neg), dtype=np.int32)
# Fix the trivial collision: neg == pos (rare but easy to handle vectorially)
for k in range(n_neg):
collision = neg_samples[:, k] == pos_idxs
neg_samples[collision, k] = (pos_idxs[collision] + 1) % self.num_movies
# Flatten: each positive gets n_neg rows
uid_rep = np.repeat(user_idxs, n_neg)
pos_rep = np.repeat(pos_idxs, n_neg)
neg_flat = neg_samples.ravel()
df = pd.DataFrame({
"user_idx": uid_rep,
"pos_movie_idx": pos_rep,
"neg_movie_idx": neg_flat,
})
logger.info(f"Generated {len(df):,} training triplets")
return df
# ------------------------------------------------------------------
# Fit / transform / save
# ------------------------------------------------------------------
def fit(
self,
ratings: pd.DataFrame,
movies: pd.DataFrame,
genome_scores: Optional[pd.DataFrame] = None,
) -> "Preprocessor":
self.movie_df = movies
self._build_id_maps(ratings, movies)
encoded = self.encode_ratings(ratings)
self.item_features = self._build_item_features(movies, genome_scores)
self.user_features = self._build_user_features(ratings, encoded)
self.fitted = True
return self
def fit_transform(
self,
ratings: pd.DataFrame,
movies: pd.DataFrame,
genome_scores: Optional[pd.DataFrame] = None,
) -> pd.DataFrame:
self.fit(ratings, movies, genome_scores)
return self.encode_ratings(ratings)
def save(self, output_dir: str | Path) -> None:
output_dir = Path(output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
# Temporarily strip movie_df before pickling.
# Pandas DataFrames with list/object columns trigger a NDArrayBacked
# __setstate__ incompatibility across pandas 2.x minor versions.
# We save movie metadata as a plain CSV instead β always portable.
movie_df_backup = self.movie_df
self.movie_df = None
with open(output_dir / "preprocessor.pkl", "wb") as f:
pickle.dump(self, f)
self.movie_df = movie_df_backup # restore in-memory object
# Save movie metadata as version-safe CSV
if movie_df_backup is not None:
meta_cols = [c for c in ["movieId", "title", "genres", "year"] if c in movie_df_backup.columns]
movie_df_backup[meta_cols].to_csv(output_dir / "movie_meta.csv", index=False)
np.save(output_dir / "item_features.npy", self.item_features)
np.save(output_dir / "user_features.npy", self.user_features)
logger.info(f"Preprocessor saved to {output_dir}")
@classmethod
def load(cls, output_dir: str | Path) -> "Preprocessor":
output_dir = Path(output_dir).resolve() # always absolute β safe regardless of CWD
with open(output_dir / "preprocessor.pkl", "rb") as f:
obj = pickle.load(f)
# Reload movie_df from CSV (bypasses pandas pickle version issues)
meta_path = output_dir / "movie_meta.csv"
if meta_path.exists():
df = pd.read_csv(meta_path)
df["genre_list"] = df["genres"].apply(
lambda g: [] if str(g) == "(no genres listed)" else str(g).split("|")
)
obj.movie_df = df
else:
obj.movie_df = None
# Also reload the numpy feature matrices in case they were updated
item_path = output_dir / "item_features.npy"
user_path = output_dir / "user_features.npy"
if item_path.exists():
obj.item_features = np.load(str(item_path))
if user_path.exists():
obj.user_features = np.load(str(user_path))
return obj
# ------------------------------------------------------------------
# Convenience entry-point
# ------------------------------------------------------------------
def preprocess_all(
data_dir: str | Path,
output_dir: str | Path,
sample_frac: Optional[float] = None,
genome_n_components: int = 32,
) -> tuple["Preprocessor", dict]:
"""
Full preprocessing pipeline: load β encode β features β split β save.
Returns (preprocessor, splits) where splits = {train, val, test, negatives}.
"""
data = load_all(data_dir, sample_frac=sample_frac)
prep = Preprocessor(genome_n_components=genome_n_components)
encoded = prep.fit_transform(
data["ratings"], data["movies"], data["genome_scores"]
)
train, val, test = prep.temporal_split(encoded)
negatives = prep.generate_negatives(train, n_neg=4)
prep.save(output_dir)
return prep, {
"train": train,
"val": val,
"test": test,
"negatives": negatives,
"movies": data["movies"],
}
|