This is my Pytorch implementation for the paper:

Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua (2019). Neural Graph Collaborative Filtering, Paper in ACM DL or Paper in arXiv. In SIGIR'19, Paris, France, July 21-25, 2019.

The author's tensorflow implementation is here

The code has been tested running under Python 3.6.10. The required packages are as follows:

• pytorch == 1.5.0
• numpy == 1.18.1
• scipy == 1.4.1
• sklearn == 0.22.1
• tqdm == 4.45.0

The instruction of commands has been clearly stated in the codes (see the parser function in NGCF/utility/parser.py).

• Gowalla dataset

python NGCF_pytorch.py --dataset gowalla --regs [1e-5] --embed_size 64 --layer_size [64,64,64] --lr 0.0001 --save_flag 1 --pretrain 0 --batch_size 1024 --epoch 400 --verbose 1 --node_dropout [0.1] --mess_dropout [0.1,0.1,0.1]

• Amazon-book dataset

python NGCF_pytorch.py --dataset amazon-book --regs [1e-5] --embed_size 64 --layer_size [64,64,64] --lr 0.0005 --save_flag 1 --pretrain 0 --batch_size 1024 --epoch 200 --verbose 50 --node_dropout [0.1] --mess_dropout [0.1,0.1,0.1]

Some important arguments:

• alg_type

  • It specifies the type of graph convolutional layer.
  • Here we provide three options:
    • ngcf (by default), proposed in Neural Graph Collaborative Filtering, SIGIR2019. Usage: --alg_type ngcf.
    • gcn, proposed in Semi-Supervised Classification with Graph Convolutional Networks, ICLR2018. Usage: --alg_type gcn.
    • gcmc, propsed in Graph Convolutional Matrix Completion, KDD2018. Usage: --alg_type gcmc.

• adj_type

  • It specifies the type of laplacian matrix where each entry defines the decay factor between two connected nodes.
  • Here we provide four options:
    • ngcf (by default), where each decay factor between two connected nodes is set as 1(out degree of the node), while each node is also assigned with 1 for self-connections. Usage: --adj_type ngcf.
    • plain, where each decay factor between two connected nodes is set as 1. No self-connections are considered. Usage: --adj_type plain.
    • norm, where each decay factor bewteen two connected nodes is set as 1/(out degree of the node + self-conncetion). Usage: --adj_type norm.
    • gcmc, where each decay factor between two connected nodes is set as 1/(out degree of the node). No self-connections are considered. Usage: --adj_type gcmc.

• node_dropout

  • It indicates the node dropout ratio, which randomly blocks a particular node and discard all its outgoing messages. Usage: --node_dropout [0.1] --node_dropout_flag 1
  • Note that the arguement node_dropout_flag also needs to be set as 1, since the node dropout could lead to higher computational cost compared to message dropout.

• mess_dropout

  • It indicates the message dropout ratio, which randomly drops out the outgoing messages. Usage --mess_dropout [0.1,0.1,0.1].

Author provides two processed datasets: Gowalla and Amazon-book. You can find it at here

• train.txt

  • Train file.
  • Each line is a user with her/his positive interactions with items: userID\t a list of itemID\n.

• test.txt

  • Test file (positive instances).
  • Each line is a user with her/his positive interactions with items: userID\t a list of itemID\n.
  • Note that here we treat all unobserved interactions as the negative instances when reporting performance.

• user_list.txt

  • User file.
  • Each line is a triplet (org_id, remap_id) for one user, where org_id and remap_id represent the ID of the user in the original and our datasets, respectively.

• item_list.txt

  • Item file.
  • Each line is a triplet (org_id, remap_id) for one item, where org_id and remap_id represent the ID of the item in the original and our datasets, respectively.