Dataset Card for WANDS (Wayfair ANnotation Dataset)

Dataset Summary

WANDS (Wayfair ANnotation Dataset) is the largest and richest publicly available dataset for e-commerce product search relevance. Created by Wayfair, this dataset enables objective benchmarking and evaluation of search engines in the e-commerce domain.

The dataset contains:

• 233,448 human-annotated (query, product) relevance judgments
• 42,994 candidate products with rich metadata
• 480 unique search query strings

Published as a companion to the ECIR 2022 paper "WANDS: Dataset for Product Search Relevance Assessment" by Yan Chen, Shujian Liu, Zheng Liu, Weiyi Sun, Linas Baltrunas and Benjamin Schroeder.

Supported Tasks

• Product Search Relevance: Evaluate whether a product is relevant to a given search query
• E-commerce Information Retrieval: Train and benchmark retrieval models for product search
• Learning-to-Rank: Build ranking models for e-commerce search results

Languages

The dataset is in English.

Dataset Structure

Data Instances

Each instance represents a query-product pair with human-annotated relevance judgment:

{
  "id": 0,
  "query_id": 0,
  "product_id": 25434,
  "label": 2,
  "product_name": "21.7 '' w waiting room chair with wood frame",
  "product_class": "Waiting Room Chairs",
  "category hierarchy": "Commercial Business Furniture / Commercial Office Furniture / Office Seating / Waiting Room Chairs / Wood Waiting Room Chairs",
  "product_description": "this is a salon chair , barber chair for a hairstylist . it is cheap , classic , hydraulic pump spa equipment .",
  "product_features": "backupholsterycolor : champagne|primarymaterial : wood|...",
  "rating_count": null,
  "average_rating": null,
  "review_count": null,
  "query": "salon chair",
  "query_class": "Massage Chairs"
}

Data Fields

• id (int): Unique identifier for the query-product pair
• query_id (int): Identifier for the search query
• product_id (int): Identifier for the product
• label (int): Human-annotated relevance label
  • 2: Exact match (product is highly relevant)
  • 1: Partial match (product is somewhat relevant)
  • 0: Irrelevant (product is not relevant)
• product_name (string): Product title/name
• product_class (string): Product classification/type
• category hierarchy (string): Full category path separated by " / "
• product_description (string): Product description text
• product_features (string): Product attributes in pipe-delimited format (key:value pairs separated by "|")
• rating_count (int/null): Number of ratings the product has received
• average_rating (float/null): Average rating score
• review_count (int/null): Number of reviews
• query (string): The search query text
• query_class (string): Predicted product class for the query

Data Splits

The dataset is provided as a single split containing all 233,448 annotated query-product pairs.

Annotation Guidelines

Relevance judgments follow three levels:

• Exact: Product matches the query intent precisely
• Partial: Product is related but not a perfect match
• Irrelevant: Product does not match the query intent

Licensing Information

This dataset is released under the Apache License 2.0.

Citation Information

@inproceedings{chen2022wands,
  title={WANDS: Dataset for Product Search Relevance Assessment},
  author={Chen, Yan and Liu, Shujian and Liu, Zheng and Sun, Weiyi and Baltrunas, Linas and Schroeder, Benjamin},
  booktitle={Advances in Information Retrieval: 44th European Conference on IR Research, ECIR 2022},
  pages={},
  year={2022},
  organization={Springer}
}

Dataset Loading

Load this dataset using the Hugging Face Datasets library:

from datasets import load_dataset

dataset = load_dataset("shuttie/wands")

Additional Resources

• GitHub Repository
• Wayfair Tech Blog Post
• Paper on Papers with Code