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.
:warning: 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.txt
To run the baselines:
Download the interaction data and images.
Generate lmdb database from the images:
cd code && python generate_lmdb.py
yaml
files to run different baselines, the yaml
files are under folders IDNet
, PixelNet
,ViNet
and overall
To run IDNet, for example, run SASRec
model on one card:
python main.py --device 0 --config_file IDNet/sasrec.yaml overall/ID.yaml
Change 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.yaml
Change 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.yaml
Change 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. $\gamma$, $\beta$ and $B$ are the learning rate, weight decay and batch size respectively.
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 |
- 2024/04/18: We have added the "description" column to the item information.