Awesome

GraphPro, WWW'2024

Introduction

GraphPro is a repository containing the source code for the research paper titled "GraphPro: Graph Pre-training and Prompt Learning for Recommendation". Our work focuses on graph-based recommendation systems in dynamic settings and explores graph pre-training and prompt learning techniques.

🌟🌟 In this repository, we include the following resources to help reproducing, benchmarking and further studying:

• Implementation of GraphPro (and GraphPro as plug-in)
• Implementation of dynamic GNNs and prompt GNNs as the baselines we use
• Preprocessed datasets for dynamic graph-based recommendation: Taobao, Koubei and Amazon
• Pre-trained weights of GraphPro and LightGCN on all the datasets

Datasets

Our preprocessed datasets are provided in dataset/. User and item interactions are split into pre-training and fine-tuning snapshots.

How to Run

Dependencies

We summarize our used packages in requirements.txt. We build the environment under Python 3.10.8.

Pre-training

🌟🌟 We alrealy included the pre-trained model weights of GraphPro and LightGCN in pretrained_weights/ to assist extremely convenient using and reproducing. :) Along with the weights we also provide pre-training logs.

To pre-train the models from scratch, the main file for pre-training is pretrain.py. You can run graph pre-training on the datasets by the following commands:

# Taobao
python pretrain.py --data_path dataset/taobao --exp_name pretrain --phase pretrain --log 1 --device cuda:0 --model GraphPro --lr 1e-3 --edge_dropout 0.5 

# Koubei
python pretrain.py --data_path dataset/koubei --exp_name pretrain --phase pretrain --log 1 --device cuda:0 --model GraphPro --lr 1e-3 --edge_dropout 0.2 --hour_interval_pre 24 

# Amazon
python pretrain.py --data_path dataset/amazon --exp_name pretrain --phase pretrain --log 1 --device cuda:0 --model GraphPro --lr 1e-3 --edge_dropout 0.2 --hour_interval_pre 24 

Fine-tuning

🌟🌟 The training logs for fine-tuning are provided in saved/[dataset]/finetune/train_log.txt for your reference during reproducing and debugging. :)

To fine-tune with the pre-trained weights, the main file is finetune.py. Relevant commands are:

# Taobao
python finetune.py --data_path dataset/taobao --exp_name finetune --log 1 --device cuda:0 --pre_model_path pretrained_weights/GraphPro_Taobao_pretrained.pt --pre_model GraphPro --f_model GraphPro --lr 1e-3 --edge_dropout 0.5 --samp_decay -0.05 

# Koubei
python finetune.py --data_path dataset/koubei --exp_name finetune --log 1 --device cuda:0 --pre_model_path pretrained_weights/GraphPro_Koubei_pretrained.pt --pre_model GraphPro --f_model GraphPro --lr 1e-3 --edge_dropout 0.2 --hour_interval_pre 24 --hour_interval_f 24 --updt_inter 2 --samp_decay 0.1

# Amazon
python finetune.py --data_path dataset/amazon --exp_name finetune --log 1 --device cuda:0 --pre_model_path pretrained_weights/GraphPro_Amazon_pretrained.pt --pre_model GraphPro --f_model GraphPro --lr 1e-3 --edge_dropout 0.2 --hour_interval_pre 24 --hour_interval_f 24 --updt_inter 4 --samp_decay -0.1 

As a plug-in on other models

Please find how we effectively implement GraphPro on other models in modules\plugins. To run relevant experiments, simply specify --plugin and change --pre_model [your_model] --f_model [your_model] in your command arguments when running finetune.py, and the script would automatically import the implemented model with GraphPro as plug-in by the following codes:

if args.plugin:
    modules_class = "modules.plugins."

For example, to run SimGCL with GraphPro, the command is:

python finetune.py --data_path dataset/taobao --exp_name finetune --log 1 --device cuda:0 --pre_model_path [your_pretrained_weights] --pre_model SimGCL --f_model SimGCL --lr 1e-3 --edge_dropout 0.5 --samp_decay -0.05 --plugin

Run Baselines -- Dynamic GNNs and Prompt GNNs

We have included our implementation of dynamic GNNs and prompt GNNs in modules\dynamicGNN and modules\graphprompt. The main files for the baselines are finetune_dynamic.py and finetune_graphprompt.py. We give two examples on how to reproduce the results for the baseline models:

python finetune_dynamic.py --data_path dataset/taobao --exp_name roland --device cuda:0 --log 1 --pre_model_path pretrained_weights/LightGCN_Taobao_pretrained.pt --f_model roland 

python finetune_graphprompt.py --data_path dataset/taobao --exp_name graphprompt --device cuda:0 --log 1 --pre_model_path pretrained_weights/LightGCN_Taobao_pretrained.pt --f_model graphprompt 

Cite

Please kindly cite our paper if you find our work, codes or data useful:

@article{yang2023graph,
  title={GraphPro: Graph Pre-training and Prompt Learning for Recommendation},
  author={Yang, Yuhao and Xia, Lianghao and Luo, Da and Lin, Kangyi and Huang, Chao},
  journal={arXiv preprint arXiv:2311.16716},
  year={2023}
}