Amazon Product Reviews with 13,000 Category Labels

AmazonCat-13k contains product descriptions and reviews classified into 13,000 multi-label categories. The dataset is split into 1.2 million training and 300,000 evaluation instances. It was originally released by McAuley and Leskovec in 2013 and was repackaged by pietrolesci in 2024 for active learning research.

Use Cases

• Train multi-label classifiers to predict product categories from review text and descriptions.
• Benchmark active learning algorithms on large-scale, imbalanced text data with thousands of labels.
• Analyze the relationship between product review content and the hierarchical category system.
• Develop models to automatically tag e-commerce products using the 13,000-category schema.

Strengths

• 1.5 million total text instances (1.2M train, 300k test)
• Extremely large label space of 13,000 categories
• Standardized split for benchmarking

Limitations

• Original data from 2013, potentially outdated for current product trends
• Inherent class imbalance common in extreme classification tasks
• No column-level metadata provided for review text and description fields

Provenance

Source

McAuley and Leskovec (2013), repackaged by pietrolesci.

Collection Method

Product descriptions and reviews from Amazon, classified into categories.

Time Range

null

Freshness

Dataset page last updated 2025-04-09; underlying data from 2013.

Geography

Presumably United States (Regionus tag), but not explicitly confirmed.