CML-HG

• The implementation of CML-HG.

Requirements

• Python == 3.6.11
• Pytorch == 1.6.0
• Numpy == 1.19.1
• Scipy == 1.5.2
• networkx == 2.5
• torch-geometric == 1.6.3
• geoopt == 0.3.1

Datasets

• We provide four processed datasets:

  • IMDB
  • ACM
  • Amazon
  • DBLP : DBLP-Citation-network V8 dataset from https://www.aminer.cn/citation
  • You can download all the preprocessed datasets used in the paper from here

• Data format:

  • Meta-paths
    • IMDB: MDM, MAM
    • ACM: PAP,PLP
    • Amazon: IVI,IBI,ITI,IOI
    • DBLP: PAP, PPP, PATAP
  • train_idx: training index, val_idx: validation index, test_idx: test index, feature: feature matrix, label: labels

How to Run

cd CML-HG
mkdir saved_model

• For running on IMDB:

python train.py --embedder CMVHG --dataset imdb --lr 0.001 --l2_coef 0.0005 --reg_coef 0.001 --w_rel 0.1 --dropadj_1 0 --dropadj_2 0 --dropfeat_1 0 --dropfeat_2 0 --sample_size 1000 --gpu 0

• For running on ACM:

python train.py --embedder CMVHG --dataset acm --lr 0.001 --l2_coef 0.0001 --reg_coef 1.0 --w_rel 0.01 --w_node 0.01 --dropadj_1 0.1 --dropadj_2 0.2 --dropfeat_1 0.1 --dropfeat_2 0.1 --isAttn --gpu 0

• For running on Amazon:

python train.py --embedder CMVHG --dataset amazon --lr 0.001 --l2_coef 0.0001 --reg_coef 0.01 --w_node 0.001 --dropadj_1 0.4 --dropadj_2 0.4 --dropfeat_1 0.1 --dropfeat_2 0.1 --sample_size 1000 --isAttn --gpu 0

• For running on DBLP:

python train.py --embedder CMVHG --dataset dblp --lr 0.001 --l2_coef 0.0005 --reg_coef 0.001 --w_rel 0.01 --w_node 0.001 --dropadj_1 0 --dropadj_2 0 --dropfeat_1 0.1 --dropfeat_2 0.1 --isAttn --gpu 0