Enhancing Graph Collaborative Filtering via Neighborhood Structure Embedding

Xinzhou Jin, Jintang Li, Yuanzhen Xie, Liang Chen, Beibei Kong, Lei Cheng, Bo Hu, Zang Li, Zibin Zheng

This is the official PyTorch implementation of our paper that has been accepted to 2023 IEEE International Conference on Data Mining (ICDM 2023). [Paper]

Requirements:

python>=3.9.13
pytorch>=1.12.1
torch-geometric>=2.2.0
torch-sparse>=0.6.15+pt112cu116
numpy>=1.24.3
pandas>=1.5.0
CUDA 11.6

Installation

# create new conda environment
conda create -n NSE python=3.9
conda activate NSE
# install some packages
pip install -r requirements.txt
# PyTorch and PyG depend on your own cuda environment. The following instructions are for CUDA11.6.
pip install torch==1.12.1+cu116 -f https://download.pytorch.org/whl/torch/
pip install torch-sparse==0.6.16 -f https://pytorch-geometric.com/whl/torch-1.12.1+cu116.html
pip install torch-scatter==2.1.0 -f https://pytorch-geometric.com/whl/torch-1.12.1+cu116.html
pip install torch-geometric

Datasets

Datasets	#Users	#Items	#Interactions	Density
ML-1M	6,039	3,628	836,478	0.03818
Yelp	45,477	30,708	1,777,765	0.00127
Books	58,144	58,051	2,517,437	0.00075
Gowalla	29,858	40,988	1,027,464	0.00084
Alibaba	300,000	81,614	1,607,813	0.00007

For ml-1m , yelp, amazon-books, gowalla-merged, they will be automatically downloaded via RecBole once you run the main program.

For alibaba, we provide it under dataset/

cd dataset
unzip alibaba.zip

Implementation of NSE:

NSE is easy to implement as follows (PyTorch-style):

def get_neighbor_adj(self):
    from torch_sparse import SparseTensor
    sp_adj = SparseTensor(row=self._user, col=self._item, value=torch.ones(
        len(self._user)), sparse_sizes=(self.n_users, self.n_items))
    return sp_adj


def get_ego_embeddings(self):
        r""" Get the embedding of users and items and combine to an embedding matrix.
        Returns:
            Tensor of the embedding matrix. Shape of [n_items+n_users, embedding_dim]
        """
        from torch_geometric.utils import spmm
        user_embeddings = spmm(self.sp_adj, self.item_embedding.weight).to_dense()
        item_embeddings = spmm(self.sp_adj.t(), self.user_embedding.weight).to_dense()
        ego_embeddings = torch.cat([user_embeddings, item_embeddings], dim=0)

        return ego_embeddings

Reproduction

We integrate our NSE-LGCN method into the RecBole and RecoBole-GNN framework.

ML-1M

python run_recbole_gnn.py --dataset "ml-1m" --model "NSELightGCN" --n_layers 3 --reg_weight 0.0001 --learning_rate 0.0002

Yelp

python run_recbole_gnn.py --dataset "yelp" --model "NSELightGCN" --n_layers 3 --reg_weight 0.01 --learning_rate 0.0001

Amazon-books

python run_recbole_gnn.py --dataset "amazon-books" --model "NSELightGCN" --n_layers 3 --reg_weight 0.001 --learning_rate 0.0001

Gowalla

python run_recbole_gnn.py --dataset "gowalla-merged" --model "NSELightGCN" --n_layers 3 --reg_weight 0.0001 --learning_rate 0.0001

Alibaba

python run_recbole_gnn.py --dataset "alibaba" --model "NSELightGCN" --n_layers 3 --reg_weight 1e-06 --learning_rate 0.0001

If you want to run on the synthetic datasets, add --ptb_strategy=replace to the above commands. For example:

python run_recbole_gnn.py --dataset "ml-1m" --model "NSELightGCN" --n_layers 3 --reg_weight 0.01 --learning_rate 0.0001 --ptb_strategy=replace

Acknowledgement

This repo is mainly based on RecBole and RecoBole-GNN. Many thanks to their wonderful work!

Citation

If you find this work is helpful to your research, please consider citing our paper:

@inproceedings{jin2023enhancing,
  title={Enhancing Graph Collaborative Filtering via Neighborhood Structure Embedding},
  author={Xinzhou Jin and Jintang Li and Yuanzhen Xie and Liang Chen and Beibei Kong and Lei Cheng and Bo Hu and Zang Li and Zibin Zheng},
  booktitle={ICDM},
  year={2023}
}

Contact

If you have any questions about this work, please feel free to contact me via [email protected]