Quick Links: 🗃️Dataset | 🛠️Experiments | 👀Others | 📭Citation | 💡News
Interaction It contains the interaction of Pixel200K, Pixel1M, Pixel8M and PixelRec, see dataset/statistics for detailed statistics.
Item Infomation It contains the item description/attributes of Pixel200K, Pixel1M, Pixel8M and PixelRec, see dataset for its detailed descriptions.
Cover It includes all the images in PixelRec, a total of 408,374 covers.
Extracted Features, currently including text feature vectors and image feature vectors.
A sampled dataset PixelRec50K was provided to help quickly understand the data contained in PixelRec. This data includes 989,494 interactions from 50,000 users with 82,865 items. The interaction data, item attributes, and covers can be downloaded here.
We provide an integrated folder for Pixel200K. After downloading the data file in this format, you can directly run the experiments in the paper under Pixel200K.
⚠️ Caution: It's prohibited to privately modify the dataset and then offer secondary downloads. If you've made alterations to the dataset in your work, you are encouraged to open-source the data processing code, so others can benefit from your methods. Or notify us of your new dataset so we can put it on this Github with your paper.
Note that this is an image recommendation dataset, if you need video information, please go to our MicroLens github https://github.com/westlake-repl/MicroLens, a large-scale micro-video recommendation dataset collected from a different platform.
Pytorch==1.10.2
cudatoolkit==11.2.1
python==3.9.7
See requirements.txt for other packages:
pip install -r requirements.txtTo run the baselines:
-
Download the interaction data and images.
-
Generate lmdb database from the images:
cd code && python generate_lmdb.py- You can choose different
yamlfiles to run different baselines, theyamlfiles are under foldersIDNet,PixelNet,ViNetandoverall
To run IDNet, for example, run SASRec model on one card:
python main.py --device 0 --config_file IDNet/sasrec.yaml overall/ID.yamlChange the IDNet/sasrec.yaml to run other IDNet baselines.
To run PixelNet, for example, run SASRec model with ViT encoders on four cards:
python main.py --device 0,1,2,3 --config_file PixelNet/sasrec.yaml overall/ViT.yamlChange PixelNet/sasrec.yaml to run other PixelNet baselines with ViT as item encoder, change overall/ViT.yaml to run sasrec model with other image encoders.
To run ViNet, e.g. run VBPR model on one card:
python main.py --device 0 --config_file ViNet/vbpr.yamlChange ViNet/vbpr.yaml to run other ViNet
Note: you may need to modify some path in files under folders ViNet and overall and file generate_lmdb.py , depending on where you put the downloaded data.
Hyper parameter range :
embedding size [128, 512, 1024, 2048, 4096, 8192]
learning rate [0.000001, 0.00005, ... , 0.001]
weight decay [0, 0.01, 0.1]
batch size [64, 128, 256, 512, 1024]
Hyper-parameter details of IDNet.
| Method (IDNet) | Model Parameters | Training Parameters |
|---|---|---|
| MF | dropout prob [0] embedding size [4096] | γ [0.0001] B [512] β [0] |
| FM | embedding size [4096] | γ [0.00005] B [64] β [0] |
| DSSM | dnn layer number [0] embedding size [4096] | γ [0.0001] B [64] β [0] |
| LightGCN | step [2] embedding size [256] | γ [0.0005] B [1024] β [0.01] |
| SASRec | trm layer number [2] inner size [2] embedding size [512] | γ [0.00005] B [64] β [0.1] |
| BERT4Rec | mask ratio [0.6] trm layer number [2] inner size [1] embedding size [512] | γ [0.00005] B [64] β [0.1] |
| LightSANs | k [3] trm layer number [1] embedding size [512] | γ [0.00005] B [512] β [0.1] |
| GRU4Rec | dropout prob [0] gru layer number [1] inner size [2] embedding size [2048] | γ [0.0001] B [64] β [0.01] |
| NextItNet | block number [3] embedding size [1024] | γ [0.0005] B [64] β [0.01] |
| SRGNN | step [2] embedding size [512] | γ [0.00005] B [64] β [0.01] |
| VisRank | visual feature [RN_2048] method [maximum] | |
| VBPR | id γ [0.001] id β [0] visual γ [0.0001] visual β [0.1] | |
| ACF | embedding size [128] | γ [0.0001] B [64] β [0.1] |
For the most architectures, PixelNet uses the same hyperparameters as its IDNet, with a few exceptions here. The embedding size refers to the hidden dimension of the user encoder.
| Method (PixelNet) | Model Parameters | Training Parameters |
|---|---|---|
| SASRec | trm layer number [2] inner size [2] embedding size [512] | γ [0.0001] B [64] β [0.1] |
| BERT4Rec | mask ratio [0.6] trm layer number [2] inner size [1] embedding size [512] | γ [0.0001] B [64] β [0.1] |
| LightSANs | k [3] trm layer number [1] embedding size [512] | γ [0.0001] B [512] β [0.1] |
| NextItNet | block number [3] embedding size [1024] | γ [0.0001] B [64] β [0.01] |
| SRGNN | step [2] embedding size [512] | γ [0.0001] B [512] β [0.01] |
In PixelNet, we adopt different learning rate and weight decay between the image encoder and the rest of the model structures. Here are the hyper-parameter for tuning the image encoders.
| Image Encoder | Hyper Parameter |
|---|---|
| RN50, RN50x4, RN50x16, RN50x64, ResNet50 | γ [0.0001] β [0.01] |
| ViT, Swin-T, Swin-B, BEiT | γ [0.0001] β [0] |
If our work has been of assistance to your work, please cite our paper as :
@article{cheng2023image,
title={An Image Dataset for Benchmarking Recommender Systems with Raw Pixels},
author={Cheng, Yu and Pan, Yunzhu and Zhang, Jiaqi and Ni, Yongxin and Sun, Aixin and Yuan, Fajie},
journal={arXiv preprint arXiv:2309.06789},
year={2023}
}
|MicroLens (A short video recommendation dataset) | https://github.com/westlake-repl/MicroLens |
|Tenrec (A dataset with 10 diverse recommendation tasks) | https://github.com/yuangh-x/2022-NIPS-Tenrec |
|NineRec (A dataset suite covering 9 downstream recommendation tasks) | https://github.com/westlake-repl/NineRec |
- 2024/11/18: We have updated a missing item "i192714" in the item information files.
- 2024/04/18: We have added the "description" column to the item information.
License
- The code in this repository is under the MIT License. Please look at the LICENSE file for details.
- See the
dataset/LICENSEfile for dataset license details.
