P5 Amazon Dataset

The P5 Amazon dataset is a large-scale product recommendation dataset containing user reviews and product metadata from Amazon.

Overview

P5 Amazon is derived from the Amazon product data and includes rich textual information about products along with user interaction histories. It's specifically designed for training generative recommendation models.

Key Features: - Multiple product categories (Beauty, Electronics, Sports, etc.) - Rich product metadata (title, brand, category, price, description) - User interaction sequences with timestamps - Pre-processed text features suitable for neural models

Dataset Structure

Data Files

When downloaded, the dataset contains:

dataset/amazon/
├── raw/
│   ├── beauty_5.json.gz          # Raw interaction data
│   ├── meta_beauty.json.gz       # Product metadata
│   └── ...
├── processed/
│   ├── items.parquet             # Processed item features
│   ├── interactions.parquet      # Processed interactions
│   └── mappings.pkl              # ID mappings
└── cache/
    └── text_embeddings/          # Cached text embeddings

Data Format

Items DataFrame: | Column | Type | Description | |--------|------|-------------| | item_id | int | Unique item identifier | | title | str | Product title | | brand | str | Product brand | | category | str | Product category | | price | float | Product price | | features | List[float] | Text embedding features (768-dim) | | text | str | Formatted text for reference |

Interactions DataFrame: | Column | Type | Description | |--------|------|-------------| | user_id | int | Unique user identifier | | item_id | int | Item identifier | | rating | float | User rating (1-5) | | timestamp | int | Interaction timestamp |

Available Categories

Beauty

• Size: ~52K products, ~1.2M interactions
• Description: Cosmetics, skincare, haircare products
• Features: Rich brand and category information

Electronics

• Size: ~63K products, ~1.7M interactions
• Description: Electronic devices, accessories, gadgets
• Features: Technical specifications in descriptions

Sports

• Size: ~35K products, ~296K interactions
• Description: Sports equipment, outdoor gear, fitness products
• Features: Activity-specific categorization

All Categories

• Total: 29 categories available
• Combined Size: Multi-GB dataset
• Use Case: Large-scale experiments

Usage

Basic Loading

from genrec.data.p5_amazon import P5AmazonItemDataset

# Load beauty category for item-level training
dataset = P5AmazonItemDataset(
    root="dataset/amazon",
    split="beauty",
    train_test_split="train"
)

print(f"Dataset size: {len(dataset)}")
print(f"Feature dimension: {dataset[0].shape}")

Sequence Data

from genrec.data.p5_amazon import P5AmazonSequenceDataset

# Load for sequence modeling (requires pre-trained RQVAE)
seq_dataset = P5AmazonSequenceDataset(
    root="dataset/amazon", 
    split="beauty",
    train_test_split="train",
    pretrained_rqvae_path="checkpoints/rqvae_beauty.ckpt"
)

# Get a sample sequence
sample = seq_dataset[0]
print(f"Input sequence: {sample['input_ids']}")
print(f"Target sequence: {sample['labels']}")

Configuration Options

from genrec.data.p5_amazon import P5AmazonItemDataset

dataset = P5AmazonItemDataset(
    root="dataset/amazon",
    split="beauty",
    train_test_split="train",

    # Text encoding options
    encoder_model_name="sentence-transformers/all-MiniLM-L6-v2",
    force_reload=False,  # Use cached embeddings if available

    # Data filtering
    min_user_interactions=5,
    min_item_interactions=5,

    # Text template
    text_template="Title: {title}; Brand: {brand}; Category: {category}"
)

Data Processing Pipeline

1. Download and Extraction

# Automatic download from UCSD repository
dataset = P5AmazonItemDataset(root="dataset/amazon", split="beauty")
# Downloads ~500MB for beauty category

2. Text Processing

The framework automatically processes product text using sentence transformers:

# Default text template
template = "Title: {title}; Brand: {brand}; Category: {category}; Price: {price}"

# Example processed text
"Title: Maybelline Mascara; Brand: Maybelline; Category: Beauty; Price: $8.99"

3. Feature Extraction

• Text Embeddings: 768-dimensional vectors from sentence transformers
• Caching: Embeddings cached for faster subsequent loading
• Normalization: L2 normalization applied by default

4. Sequence Building

For TIGER training, interactions are converted to sequences:

# User interaction history
user_history = [item1, item2, item3, item4]

# Converted to semantic ID sequence using RQVAE
semantic_sequence = [1, 45, 123, 67, 234, 189, 45, 123, 567, 234, 88, 192]
#                   |--item1--| |--item2--| |--item3--| |--item4--|

Statistics

Beauty Category

Items: 52,024
Users: 40,226  
Interactions: 1,235,316
Density: 0.059%
Avg items per user: 30.7
Avg users per item: 23.7

Electronics Category

Items: 63,001
Users: 192,403
Interactions: 1,689,188
Density: 0.014%
Avg items per user: 8.8
Avg users per item: 26.8

Data Quality

Preprocessing Steps

1. Duplicate Removal: Remove duplicate user-item interactions
2. Low Activity Filtering: Filter users/items with < 5 interactions
3. Text Cleaning: Normalize titles, handle missing brands/categories
4. Price Processing: Clean and standardize price formats
5. ID Remapping: Create contiguous ID mappings

Quality Checks

from genrec.data.p5_amazon import P5AmazonItemDataset

dataset = P5AmazonItemDataset(root="dataset/amazon", split="beauty")

# Check data quality
items_df, interactions_df = dataset.base_dataset.get_dataset()

print("Data Quality Report:")
print(f"Items with missing titles: {items_df['title'].isna().sum()}")
print(f"Items with missing brands: {items_df['brand'].isna().sum()}")
print(f"Interactions with valid ratings: {(interactions_df['rating'] > 0).sum()}")
print(f"Feature vector dimension: {len(items_df.iloc[0]['features'])}")

Advanced Usage

Custom Text Templates

# Product-focused template
template = "Product: {title} from {brand} in {category} category"

# Price-aware template  
template = "Buy {title} by {brand} for ${price} in {category}"

# Minimal template
template = "{title} - {brand}"

Batch Processing

from torch.utils.data import DataLoader

dataset = P5AmazonItemDataset(root="dataset/amazon", split="beauty")
dataloader = DataLoader(dataset, batch_size=128, shuffle=True, num_workers=4)

for batch in dataloader:
    # batch shape: (128, 768)
    features = batch
    # Process batch...

Multi-Category Loading

# Load multiple categories
categories = ["beauty", "electronics", "sports"]
datasets = []

for category in categories:
    dataset = P5AmazonItemDataset(
        root="dataset/amazon",
        split=category,
        train_test_split="train"
    )
    datasets.append(dataset)

# Combine datasets
from torch.utils.data import ConcatDataset
combined_dataset = ConcatDataset(datasets)

Troubleshooting

Common Issues

Q: Download fails with network error A: Check internet connection and try again. The files are large (100MB-2GB).

Q: Text encoding takes very long A: Use cached embeddings by setting force_reload=False and ensure cache directory is writable.

Q: Out of memory during loading A: Reduce batch size or use a smaller category like "beauty" instead of "all".

Q: Missing brand/category information A: This is normal - the dataset fills missing values with "Unknown".

Performance Tips

# Use caching for faster subsequent loads
dataset = P5AmazonItemDataset(
    root="dataset/amazon",
    split="beauty", 
    force_reload=False  # Use cache
)

# Use lighter text encoder for faster processing
dataset = P5AmazonItemDataset(
    root="dataset/amazon",
    split="beauty",
    encoder_model_name="sentence-transformers/all-MiniLM-L6-v2"  # Smaller model
)

# Process in smaller batches
from genrec.data.configs import TextEncodingConfig
text_config = TextEncodingConfig(batch_size=16)  # Reduce from default 32

Citation

If you use the P5 Amazon dataset, please cite:

@article{geng2022recommendation,
  title={Recommendation as language processing (rlp): A unified pretrain, personalized prompt \& predict paradigm (p5)},
  author={Geng, Shijie and Liu, Shuchang and Fu, Zuohui and Ge, Yingqiang and Zhang, Yongfeng},
  journal={arXiv preprint arXiv:2203.13366},
  year={2022}
}

Related Documentation

• Dataset Overview - General dataset concepts
• Custom Datasets - Creating your own datasets
• RQVAE Training - Training item encoders
• TIGER Training - Training sequence models