NARS is an algorithm for node classification on heterogeneous graphs, based on scalable neighbor averaging techniques that have been previously used in e.g. SIGN to heterogeneous scenarios by generating neighbor-averaged features on sampled relation induced subgraphs.

For more details, please check out our paper:

Scalable Graph Neural Networks for Heterogeneous Graphs

• torch==1.5.1+cu101
• dgl-cu101==0.4.3.post2
• ogb==1.2.1
• dglke==0.1.0

We have prepared a dockerfile for building a container with clean environment and all required dependencies. Please checkout instructions in docker.

If you plan to evaluate on OAG dataset, you need to follow instructions in oag_dataset to download and pre-process dataset.

In academic graph datasets (ACM, MAG, OAG) in which only paper nodes are associated with input features. NARS featurizes other node types with TransE relational graph embedding pre-trained on the graph structure.

Please follow instructions in graph_embed to generate embeddings for each dataset.

NARS samples Relation Subsets (see our paper for details). Please follow the instructions in sample_relation_subsets to generate these subsets.

Or you may skip this step and use the example subsets that have added to this repository.

NARS are evaluated on three academic graph datasets to predict publishing venues and fields of papers.

python3 train.py --dataset acm --use-emb TransE_acm --R 2 \
    --use-relation-subsets sample_relation_subsets/examples/acm \
    --num-hidden 64 --lr 0.003 --dropout 0.7 --eval-every 1 \
    --num-epochs 100 --input-dropout

python3 train.py --dataset mag --use-emb TransE_mag --R 5 \
    --use-relation-subset sample_relation_subsets/examples/mag \
    --eval-batch-size 50000 --num-hidden 512 --lr 0.001 --batch-s 50000 \
    --dropout 0.5 --num-epochs 1000

python3 train.py --dataset oag_venue --use-emb TransE_oag_venue --R 3 \
    --use-relation-subsets sample_relation_subsets/examples/oag_venue \
    --eval-batch-size 25000 --num-hidden 256 --lr 0.001 --batch-size 1000 \
    --data-dir oag_dataset --dropout 0.5 --num-epochs 200

python3 train.py --dataset oag_L1 --use-emb TransE_oag_L1 --R 3 \
    --use-relation-subsets sample_relation_subsets/examples/oag_L1 \
    --eval-batch-size 25000 --num-hidden 256 --lr 0.001 --batch-size 1000 \
    --data-dir oag_dataset --dropout 0.5 --num-epochs 200

Here is a summary of model performance using example relation subsets:

For ACM and OGBN-MAG dataset, the task is to predict paper publishing venue.

For OAG dataset, there are two different node predictions tasks: predicting venue (single-label) and L1-field (multi-label). And we follow Heterogeneous Graph Transformer to evaluate using NDCG and MRR metrics.

The above commands were tested on Tesla V100 (32 GB) and Tesla T4 (15GB). If your GPU memory isn't enough for handling large graphs, try the following:

• add --cpu-process to the command to move preprocessing logic to CPU
• reduce evaluation batch size with --eval-batch-size. The evaluation result won't be affected since model is fixed.
• reduce training batch with --batch-size

As mentioned in our paper, using a lot of relation subsets may consume too much CPU memory. To reduce CPU memory footprint, we implemented an optimization in train_partial.py which trains part of our feature aggregation weights at a time.

Using OGBN-MAG dataset as an example, the following command randomly picks 3 subsets from all 8 sampled relation subsets and trains their aggregation weights every 10 epochs.

python3 train_partial.py --dataset mag --use-emb TransE_mag --R 5 \
    --use-relation-subsets sample_relation_subsets/examples/mag \
    --eval-batch-size 50000 --num-hidden 512 --lr 0.001 --batch-size 50000 \
    --dropout 0.5 --num-epochs 1000 --sample-size 3 --resample-every 10

Please cite our paper with:

@article{yu2020scalable,
    title={Scalable Graph Neural Networks for Heterogeneous Graphs},
    author={Yu, Lingfan and Shen, Jiajun and Li, Jinyang and Lerer, Adam},
    journal={arXiv preprint arXiv:2011.09679},
    year={2020}
}

NARS is CC-by-NC licensed, as found in the LICENSE file.