[SDM2024] PixelRec: An Image Dataset for Benchmarking Recommender Systems with Raw Pixels

Quick Links: 🗃️Dataset | 🛠️Experiments | 👀Others | 📭Citation | 💡News

Dataset

Overview

Download Link

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.

Experiments

Environments

Pytorch==1.10.2
cudatoolkit==11.2.1
python==3.9.7

See requirements.txt for other packages:

pip install -r requirements.txt

Run Baselines

To 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 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 Parameters

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]

Citation

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}
}

More Resources：

|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 |

News

- 2024/04/18: We have added the "description" column to the item information.

💡 If you have an innovative idea for building a foundation recommendation model but require a large dataset and computational resources, consider joining our lab as an intern or visiting scholar. We can provide access to 100 NVIDIA 80G A100 GPUs and a billion-level dataset of user-image/text interactions.

The laboratory is hiring research assistants, interns, doctoral students, and postdoctoral researchers. Please contact the corresponding author for details.

实验室招聘科研助理，实习生，博士生和博士后，请联系通讯作者。