metadata

language:
  - en
license: mit
size_categories:
  - 100K<n<1M
pretty_name: Shopping Queries Image Dataset (SQID 🦑)
configs:
  - config_name: product_image_urls
    default: true
    data_files:
      - split: train
        path: data/product_image_urls.parquet
  - config_name: product_features
    data_files:
      - split: train
        path: data/product_features.parquet
  - config_name: query_features
    data_files:
      - split: train
        path: data/query_features.parquet
  - config_name: supp_product_image_urls
    data_files:
      - split: train
        path: data/supp_product_image_urls.parquet
dataset_info:
  - config_name: product_image_urls
    features:
      - name: product_id
        dtype: string
      - name: image_url
        dtype: string
  - config_name: product_features
    features:
      - name: product_id
        dtype: string
      - name: clip_text_features
        sequence: float32
      - name: clip_image_features
        sequence: float32
  - config_name: query_features
    features:
      - name: query_id
        dtype: string
      - name: clip_text_features
        sequence: float32
  - config_name: supp_product_image_urls
    features:
      - name: product_id
        dtype: string
      - name: image_url
        dtype: string

Shopping Queries Image Dataset (SQID 🦑): An Image-Enriched ESCI Dataset for Exploring Multimodal Learning in Product Search

Introduction

The Shopping Queries Image Dataset (SQID) is a dataset that includes image information for over 190,000 products. This dataset is an augmented version of the Amazon Shopping Queries Dataset, which includes a large number of product search queries from real Amazon users, along with a list of up to 40 potentially relevant results and judgments of how relevant they are to the search query.

The image-enriched SQID dataset can be used to support research on improving product search by leveraging image information. Researchers can use this dataset to train multimodal machine learning models that can take into account both textual and visual information when ranking products for a given search query.

Dataset

This dataset extends the Shopping Queries Dataset (SQD) by including image information and visual embeddings for each product, as well as text embeddings for the associated queries which can be used for baseline product ranking benchmarking.

Product Sampling

We limited this dataset to the subset of the SQD where small_version is 1 (the reduced version of the dataset for Task 1), split is 'test' (test set of the dataset), and product_locale is 'us'. Hence, this dataset includes 164,900 product_id's

As supplementary data, we also provide data related to the other products appearing in at least 2 query judgements in the data of Task 1 with product_locale as 'us', amounting to 27,139 products, to further increase the coverage of the data for additional applications that go beyond the ESCI benchmark.

Image URL Scraping:

We scraped 156,545 (95% of the 164,900 product_id's) image_urls from the Amazon website. Products lacking image_urls either failed to fetch a valid product page (usually if Amazon no longer sells the product) or displayed a default "No image available" image.

Note: 446 product image_urls are a default digital video image, 'https://m.media-amazon.com/images/G/01/digital/video/web/Default_Background_Art_LTR._SX1080_FMjpg_.jpg', implying no product-specific image exists.

The dataset also includes a supplementary file covering 27,139 more product_id's and image_url's.

Image Embeddings:

We extracted image embeddings for each of the images using the OpenAI CLIP model from HuggingFace, specifically clip-vit-large-patch14, with all default settings.

Query Embeddings:

For each query and each product in the SQD Test Set, we extracted text embeddings using the same CLIP model and based on the query text and product title. These can be useful to benchmark a baseline product search method where both text and images share the same embedding space.

Files

The data directory contains 4 files:

product_image_urls.parquet
- This file contains the image URLs for all product_id's in the dataset
products_features.parquet
- This file contains the CLIP embedding features for all product_id's in the dataset
queries_features.parquet
- This file contains the CLIP text embedding features for all querie_id's in the dataset
supp_product_image_urls.parquet
- This file contains supplementary data as image URLs for an additional set of products not included in the test set and increasing the coverage of the data

Code snippets to get CLIP features

SQID includes embeddings extracted using OpenAI CLIP model from HuggingFace (clip-vit-large-patch14). We provide below code snippets in Python to extract such embeddings, using either the model from HuggingFace or using Replicate.

Using CLIP model from HuggingFace

from PIL import Image
import requests
from transformers import CLIPModel, CLIPProcessor

model = CLIPModel.from_pretrained("openai/clip-vit-large-patch14")
processor = CLIPProcessor.from_pretrained("openai/clip-vit-large-patch14")

image = Image.open(requests.get('https://m.media-amazon.com/images/I/71fv4Dv5RaL._AC_SY879_.jpg', stream=True).raw)
inputs = processor(images=[image], return_tensors="pt", padding=True)
image_embds = model.get_image_features(pixel_values=inputs["pixel_values"])

Using Replicate

import replicate

client = replicate.Client(api_token=REPLICATE_API_KEY)
output = client.run(
    "andreasjansson/clip-features:71addf5a5e7c400e091f33ef8ae1c40d72a25966897d05ebe36a7edb06a86a2c",
    input={
        "inputs": 'https://m.media-amazon.com/images/I/71fv4Dv5RaL._AC_SY879_.jpg'
    }
)

Citation

To use this dataset, please cite the following paper:

Shopping Queries Image Dataset (SQID): An Image-Enriched ESCI Dataset for Exploring Multimodal Learning in Product Search, M. Al Ghossein, C.W. Chen, J. Tang

License

This dataset is released under the MIT License

Acknowledgments

SQID was developed at Crossing Minds by:

This dataset would not have been possible without the amazing Shopping Queries Dataset by Amazon.