HGCN: A Heterogeneous Graph Convolutional Network-Based Deep Learning Model Toward Collective Classification

• Python 2.7
• TensorFlow 1.14.0

Here, we provide two real-world HIN datasets: CORA and IMDB.

Run HGCN training on the CORA dataset:

$ python train.py --dataset cora --kernel-size 4 --inception-depth 1 --label-propagation 0 --epochs 30

Run HGCN training on the IMDB dataset:

$ python train.py --dataset imdb --kernel-size 2 --inception-depth 1 --label-propagation 0 --epochs 30

If you want to train HGCN on your own dataset, you should prepare the following four files:

• *.adj.npz: The adjacency matrix for each type of edges.
• *.feat.label.npz: The one-hot codes of the labels of target-type nodes. Note that, 0 to initialize the features of nontarget-type nodes.
• *.label.all: The labels of all target-type nodes. Each line contains one token <label>.
• *.label.part: the target-type nodes that have the labels, and their labels. Each line contains two token <node> <label>.