reczoo/Criteo_x4

Criteo_x4

• —Dataset description:

The Criteo dataset is a widely-used benchmark dataset for CTR prediction, which contains about one week of click-through data for display advertising. It has 13 numerical feature fields and 26 categorical feature fields. Following the setting with the AutoInt work, we randomly split the data into 8:1:1 as the training set, validation set, and test set, respectively.

The dataset statistics are summarized as follows:

• —Criteox4001

In this setting, we follow the winner's solution of the Criteo challenge to discretize each integer value x to ⌊log2(x)⌋, if x > 2; and x = 1 otherwise. For all categorical fields, we replace infrequent features with a default `<OOV>` token by setting the threshold mincategorycount=10. Note that we do not follow the exact preprocessing steps in AutoInt, because this preprocessing performs much better. We fix embedding_dim=16 as with AutoInt.

• —Criteox4002

In this setting, we follow the winner's solution of the Criteo challenge to discretize each integer value x to ⌊log2(x)⌋, if x > 2; and x = 1 otherwise. For all categorical fields, we replace infrequent features with a default `<OOV>` token by setting the threshold mincategorycount=2. We fix embedding_dim=40 in this setting.

• —Source: https://www.kaggle.com/c/criteo-display-ad-challenge/data
• —Download: https://huggingface.co/datasets/reczoo/Criteo_x4/tree/main
• —RecZoo Datasets: https://github.com/reczoo/Datasets

• —Used by papers:
• —Weiping Song, Chence Shi, Zhiping Xiao, Zhijian Duan, Yewen Xu, Ming Zhang, Jian Tang. AutoInt: Automatic Feature Interaction Learning via Self-Attentive Neural Networks. In CIKM 2019.
• —Jieming Zhu, Jinyang Liu, Shuai Yang, Qi Zhang, Xiuqiang He. BARS-CTR: Open Benchmarking for Click-Through Rate Prediction. In CIKM 2021.

• —Check the md5sum for data integrity:

  $ md5sum train.csv valid.csv test.csv
  4a53bb7cbc0e4ee25f9d6a73ed824b1a  train.csv
  fba5428b22895016e790e2dec623cb56  valid.csv
  cfc37da0d75c4d2d8778e76997df2976  test.csv