KuaiRec is a real-world dataset collected from the recommendation logs of the video-sharing mobile app Kuaishou. For now, it is the first dataset that contains a fully observed user-item interaction matrix. For the term "fully observed", we mean there are almost no missing values in the user-item matrix, i.e., each user has viewed each video and then left feedback.

Other related open-sourced datasets are KuaiRand and KuaiSAR.

The following figure illustrates the user-item matrices in traditional datasets and KuaiRec.

With all user preferences known, KuaiRec can used in offline evaluation (i.e., offline A/B test) for recommendation models. It can benefit lots of research directions, such as unbiased recommendation, interactive/conversational recommendation, reinforcement learning (RL), and off-policy evaluation (OPE) for recommendation.

If you use it in your work, please cite our paper:

@inproceedings{gao2022kuairec,
  author = {Gao, Chongming and Li, Shijun and Lei, Wenqiang and Chen, Jiawei and Li, Biao and Jiang, Peng and He, Xiangnan and Mao, Jiaxin and Chua, Tat-Seng},
  title = {KuaiRec: A Fully-Observed Dataset and Insights for Evaluating Recommender Systems},
  booktitle = {Proceedings of the 31st ACM International Conference on Information \& Knowledge Management},
  series = {CIKM '22},
  location = {Atlanta, GA, USA},
  url = {https://doi.org/10.1145/3511808.3557220},
  doi = {10.1145/3511808.3557220},
  numpages = {11},
  year = {2022},
  pages = {540–550}
}

This repository lists the example codes in evaluating conversational recommendations as described in the paper.

We provide some simple statistics of this dataset here. It is generated by Statistics_KuaiRec.ipynb. You can do it online at Google Colab .

2022.05.16: We update the dataset to version 2.0. We made the following changes:

• We removed the unused video ID=1225 from all tables having the field video_id and re-indexed the rest videos, i.e., ID = ID - 1 if ID > 1225.
• We added two tables to enhance the side information for users and videos, respectively. See 4.item_daily_features.csv and 5. user_features.csv under the data description section for details.

We provide several options to download this dataset:

Option 1. Download via the "wget" command.

 wget https://nas.chongminggao.top:4430/datasets/KuaiRec.zip --no-check-certificate
 unzip KuaiRec.zip

Option 2. Download manually through the following links:

• Optional link 1: Google Drive

• Optional link 2: USTC Drive (中科大)

The script loaddata.py provides a simple way to load the data via Pandas in Python.

KuaiRec contains millions of user-item interactions as well as side information including the item categories and a social network. Six files are included in the download data:

KuaiRec
├── data
│   ├── big_matrix.csv          
│   ├── small_matrix.csv
│   ├── social_network.csv
│   ├── user_features.csv
│   ├── item_daily_features.csv
│   └── item_categories.csv

The statistics of the small matrix and big matrix in KuaiRec.

Note that the density of the small matrix is 99.6% instead of 100% because some users have explicitly indicated that they would not be willing to receive recommendations from certain authors. I.e., They blocked these videos.

The "watch_ratio" can be deemed as the label of the interaction. Note: there is no "like" signal for this dataset. If you need this binary signal in your scenarios, you can create it yourself. E.g., like = 1 if watch_ratio > 2.0.