File size: 13,462 Bytes
2319f81 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 | """
Data pipeline for MARS sequential recommendation.
Supports:
1. Amazon Reviews 2023 (Books, Movies_and_TV, etc.) — filtered for power users
2. MovieLens-1M
3. Synthetic data for testing
All data is converted to a unified format:
- user_id: int
- item_ids: List[int] (chronologically sorted)
- timestamps: List[float] (in seconds)
"""
import os
import random
import numpy as np
import torch
from torch.utils.data import Dataset, DataLoader
from typing import Dict, List, Tuple, Optional
from collections import defaultdict
import json
def download_movielens_1m(data_dir: str = './data/ml-1m') -> str:
"""Download and extract MovieLens-1M dataset."""
import urllib.request
import zipfile
os.makedirs(data_dir, exist_ok=True)
ratings_path = os.path.join(data_dir, 'ratings.dat')
if not os.path.exists(ratings_path):
url = 'https://files.grouplens.org/datasets/movielens/ml-1m.zip'
zip_path = os.path.join(data_dir, 'ml-1m.zip')
print(f"Downloading MovieLens-1M from {url}...")
urllib.request.urlretrieve(url, zip_path)
with zipfile.ZipFile(zip_path, 'r') as z:
z.extractall(data_dir)
# Move files up one level
inner_dir = os.path.join(data_dir, 'ml-1m')
if os.path.exists(inner_dir):
for f in os.listdir(inner_dir):
os.rename(os.path.join(inner_dir, f), os.path.join(data_dir, f))
os.rmdir(inner_dir)
os.remove(zip_path)
return ratings_path
def load_movielens_1m(data_dir: str = './data/ml-1m', min_interactions: int = 5):
"""Load MovieLens-1M and return user sequences."""
ratings_path = download_movielens_1m(data_dir)
# Parse ratings.dat
user_interactions = defaultdict(list)
with open(ratings_path, 'r') as f:
for line in f:
parts = line.strip().split('::')
user_id = int(parts[0])
item_id = int(parts[1])
rating = float(parts[2])
timestamp = int(parts[3])
# Keep all ratings (implicit feedback style)
user_interactions[user_id].append((item_id, timestamp))
# Sort by timestamp, filter short sequences
sequences = {}
for uid, interactions in user_interactions.items():
interactions.sort(key=lambda x: x[1])
if len(interactions) >= min_interactions:
sequences[uid] = {
'item_ids': [x[0] for x in interactions],
'timestamps': [float(x[1]) for x in interactions]
}
return sequences
def load_amazon_reviews(
category: str = 'Movies_and_TV',
min_interactions: int = 20,
max_users: int = 50000,
data_dir: str = './data/amazon'
):
"""
Load Amazon Reviews 2023 dataset from HF Hub.
Filters to users with min_interactions+ for long-sequence modeling.
"""
try:
from datasets import load_dataset
print(f"Loading Amazon Reviews 2023 - {category}...")
# Try benchmark format first
try:
ds = load_dataset(
"McAuley-Lab/Amazon-Reviews-2023",
f"0core_rating_only_{category}",
trust_remote_code=True,
split="train"
)
except Exception:
# Fallback to raw reviews
ds = load_dataset(
"McAuley-Lab/Amazon-Reviews-2023",
f"raw_review_{category}",
trust_remote_code=True,
split="full"
)
# Build user sequences
user_interactions = defaultdict(list)
for row in ds:
uid = row.get('user_id', row.get('reviewerID'))
iid = row.get('parent_asin', row.get('asin'))
ts = row.get('timestamp', row.get('unixReviewTime', 0))
if uid and iid:
user_interactions[uid].append((iid, float(ts) / 1000 if ts > 1e12 else float(ts)))
# Filter and sort
sequences = {}
for uid, interactions in user_interactions.items():
interactions.sort(key=lambda x: x[1])
if len(interactions) >= min_interactions:
sequences[uid] = {
'item_ids': [x[0] for x in interactions],
'timestamps': [x[1] for x in interactions]
}
# Limit users
if len(sequences) > max_users:
keys = random.sample(list(sequences.keys()), max_users)
sequences = {k: sequences[k] for k in keys}
return sequences
except Exception as e:
print(f"Failed to load Amazon Reviews: {e}")
return {}
def generate_synthetic_data(
num_users: int = 5000,
num_items: int = 10000,
min_seq_len: int = 50,
max_seq_len: int = 1000,
seed: int = 42
) -> Dict:
"""
Generate synthetic sequential interaction data for testing.
Simulates realistic patterns:
- Power law item popularity
- Temporal patterns (daily/weekly)
- User interest drift over time
"""
rng = np.random.RandomState(seed)
# Power law item popularity
item_popularity = rng.power(0.8, num_items)
item_popularity /= item_popularity.sum()
sequences = {}
base_time = 1600000000 # ~Sep 2020
for uid in range(num_users):
seq_len = rng.randint(min_seq_len, max_seq_len + 1)
# User has a few interest clusters
num_clusters = rng.randint(2, 6)
cluster_centers = rng.choice(num_items, num_clusters, replace=False)
cluster_weights = rng.dirichlet(np.ones(num_clusters))
items = []
timestamps = []
current_time = base_time + rng.randint(0, 86400 * 365) # Random start
for t in range(seq_len):
# Interest drift: cluster weights shift over time
drift = rng.dirichlet(np.ones(num_clusters) * 5)
current_weights = 0.8 * cluster_weights + 0.2 * drift
# Select cluster, then item near cluster center
cluster = rng.choice(num_clusters, p=current_weights / current_weights.sum())
center = cluster_centers[cluster]
# Items near cluster center (with some randomness)
local_items = np.arange(
max(0, center - 50),
min(num_items, center + 50)
)
local_probs = item_popularity[local_items]
local_probs /= local_probs.sum()
item = local_items[rng.choice(len(local_items), p=local_probs)]
items.append(int(item) + 1) # 1-indexed (0 = padding)
# Time gap: exponential with daily/weekly patterns
gap = rng.exponential(3600) # avg 1 hour
# Add daily pattern
hour = (current_time % 86400) / 3600
if 2 < hour < 8: # Less activity at night
gap *= 3
current_time += gap
timestamps.append(current_time)
sequences[uid] = {
'item_ids': items,
'timestamps': timestamps
}
return sequences
class ReindexedData:
"""Reindex items to contiguous integers and provide train/val/test splits."""
def __init__(
self,
sequences: Dict,
max_seq_len: int = 512,
val_ratio: float = 0.1,
test_ratio: float = 0.1,
):
self.max_seq_len = max_seq_len
# Collect all items and reindex
all_items = set()
for uid, data in sequences.items():
all_items.update(data['item_ids'])
self.item2idx = {item: idx + 1 for idx, item in enumerate(sorted(all_items))}
self.idx2item = {idx: item for item, idx in self.item2idx.items()}
self.num_items = len(self.item2idx)
print(f"Total users: {len(sequences)}, Total items: {self.num_items}")
# Reindex and split
self.train_data = []
self.val_data = []
self.test_data = []
seq_lens = []
for uid, data in sequences.items():
item_ids = [self.item2idx[i] for i in data['item_ids']]
timestamps = data['timestamps']
# Truncate to max_seq_len
if len(item_ids) > max_seq_len:
item_ids = item_ids[-max_seq_len:]
timestamps = timestamps[-max_seq_len:]
seq_lens.append(len(item_ids))
if len(item_ids) < 3:
continue
# Leave-one-out split
self.train_data.append({
'user_id': uid,
'item_ids': item_ids[:-2],
'timestamps': timestamps[:-2],
'next_item': item_ids[-2],
})
self.val_data.append({
'user_id': uid,
'item_ids': item_ids[:-1],
'timestamps': timestamps[:-1],
'next_item': item_ids[-1],
})
self.test_data.append({
'user_id': uid,
'item_ids': item_ids[:-1],
'timestamps': timestamps[:-1],
'next_item': item_ids[-1],
})
seq_lens = np.array(seq_lens)
print(f"Sequence length stats: mean={seq_lens.mean():.1f}, "
f"median={np.median(seq_lens):.1f}, "
f"max={seq_lens.max()}, min={seq_lens.min()}")
print(f"Users with 100+ interactions: {(seq_lens >= 100).sum()}")
print(f"Users with 200+ interactions: {(seq_lens >= 200).sum()}")
print(f"Train: {len(self.train_data)}, Val: {len(self.val_data)}, "
f"Test: {len(self.test_data)}")
class SeqRecDataset(Dataset):
"""Sequential recommendation dataset with negative sampling."""
def __init__(
self,
data: List[Dict],
num_items: int,
max_seq_len: int = 512,
num_negatives: int = 1,
is_training: bool = True,
):
self.data = data
self.num_items = num_items
self.max_seq_len = max_seq_len
self.num_negatives = num_negatives
self.is_training = is_training
def __len__(self):
return len(self.data)
def __getitem__(self, idx):
sample = self.data[idx]
item_ids = sample['item_ids'][-self.max_seq_len:]
timestamps = sample['timestamps'][-self.max_seq_len:]
next_item = sample['next_item']
# Padding
seq_len = len(item_ids)
pad_len = self.max_seq_len - seq_len
# Right-padding (needed for causal attention to work correctly)
padded_items = item_ids + [0] * pad_len
padded_timestamps = timestamps + [0.0] * pad_len
mask = [True] * seq_len + [False] * pad_len
# Negative sampling
item_set = set(item_ids)
negatives = []
for _ in range(self.num_negatives):
neg = random.randint(1, self.num_items)
while neg in item_set:
neg = random.randint(1, self.num_items)
negatives.append(neg)
return {
'item_ids': torch.tensor(padded_items, dtype=torch.long),
'timestamps': torch.tensor(padded_timestamps, dtype=torch.float32),
'mask': torch.tensor(mask, dtype=torch.bool),
'positive_ids': torch.tensor(next_item, dtype=torch.long),
'negative_ids': torch.tensor(negatives, dtype=torch.long),
}
def create_dataloaders(
data: ReindexedData,
max_seq_len: int = 512,
batch_size: int = 128,
num_negatives: int = 4,
eval_negatives: int = 999,
num_workers: int = 2,
) -> Tuple[DataLoader, DataLoader, DataLoader]:
"""Create train/val/test dataloaders.
Uses 999 negatives for evaluation (standard SASRec protocol).
"""
train_dataset = SeqRecDataset(
data.train_data, data.num_items, max_seq_len,
num_negatives=num_negatives, is_training=True
)
val_dataset = SeqRecDataset(
data.val_data, data.num_items, max_seq_len,
num_negatives=eval_negatives, is_training=False
)
test_dataset = SeqRecDataset(
data.test_data, data.num_items, max_seq_len,
num_negatives=eval_negatives, is_training=False
)
train_loader = DataLoader(
train_dataset, batch_size=batch_size, shuffle=True,
num_workers=num_workers, pin_memory=True, drop_last=True,
)
val_loader = DataLoader(
val_dataset, batch_size=batch_size, shuffle=False,
num_workers=num_workers, pin_memory=True,
)
test_loader = DataLoader(
test_dataset, batch_size=batch_size, shuffle=False,
num_workers=num_workers, pin_memory=True,
)
return train_loader, val_loader, test_loader
def save_data_config(data: ReindexedData, path: str):
"""Save data configuration for model loading."""
config = {
'num_items': data.num_items,
'num_train': len(data.train_data),
'num_val': len(data.val_data),
'num_test': len(data.test_data),
}
os.makedirs(os.path.dirname(path) if os.path.dirname(path) else '.', exist_ok=True)
with open(path, 'w') as f:
json.dump(config, f, indent=2)
return config
|