The official implementation for "NodeFormer: A Scalable Graph Structure Learning Transformer for Node Classification" which is accepted to NeurIPS22 as a spotlight presentation.

Related materials: [paper], [slides], [blog English | Chinese], [video English | Chinese]

This work takes an initial step for exploring Transformer-style graph encoder networks for large node classification graphs, dubbed as NodeFormer, as an alternative to common Graph Neural Networks, in particular for encoding nodes in an input graph into embeddings in latent space.

NodeFormer is a pioneering Transformer model for node classification on large graphs. NodeFormer scales all-pair message passing with efficient latent structure learning to million-level nodes. NodeFormer has several merits:

• All-Pair Message Passing on Layer-specific Adaptive Structures. The feature propagation per layer is operated on a latent graph that potentially connect all the nodes, in contrast with the local propagation design of GNNs that only aggregates the embeddings of neighbored nodes.

• Linear Complexity w.r.t. Node Numbers. The all-pair message passing on latent graphs that are optimized together only requires $O(N)$ complexity, empowered by our proposed kernelized Gumbel-Softmax message passing. The largest demonstration of our model in our paper is the graph with 2M nodes, yet we believe it can even scale to larger ones with the mini-batch partition.

• Efficient End-to-End Learning for Latent Structures. The optimization for the latent structures is allowed for end-to-end training with the model, making the whole learning process simple and efficient. E.g., the training on Cora only requires 1-2 minutes, while on OGBN-Proteins requires 1-2 hours in one run.

• Flexibility for Inductive Learning and Graph-Free Scenarios. NodeFormer is flexible for handling new unseen nodes in testing and as well as predictive tasks without input graphs, e.g., image and text classification. It can also be used for interpretability analysis with the latent interactions among data points explicitly estimated.

The following figure presents how we achieve $O(N)$ complexity by means of the kernelized Gumbel-Softmax message passing that organically combines Random Feature Map and Gumbel-Softmax strategies.

The data flow of NodeFormer is shown by the following figure. In each layer, it consists of three components: all-pair message passing by the kernelized Gumbel-Softmax (red box), adding relational bias (green box) and computing edge regularization loss (blue box).

The nodeformer.py implements our model:

• The functions kernelized_softmax() and kernelized_gumbel_softmax() implement the message passing per layer. The Gumbel version is only used for training.

• The layer class NodeFormerConv implements one-layer feed-forward of NodeFormer which contains three operations:

  • Feature propagation on a latent (all-pair) graph using kernelized (Gumbel-)Softmax
  • (optional) Adding relational bias by input graph adjacency (using 1 or 2 hop)
  • (optional) Computing edge-level regularization loss from observed edges in the input graph

• The model class NodeFormer implements the model that adopts standard input (node features, adjacency) and output (node prediction, edge loss) where edge loss is optional.

For other files, the descriptions are below:

• main.py is the pipeline for full-graph training/evaluation.

• main-batch.py is the pipeline for training with random mini-batch partition for large datasets.

• parse.py for hyper-parameter arguments, logger.py for result recording and printing, eval.py for model evaluation, dataset.py for loader and preprocessing of each dataset, data_utils contains some functions for data analysis/processing, gnn.py implements common GNNs.

The datasets we used span three categories (Sec 4.1, 4.2, 4.3 in our paper)

• Transductive Node Classification: we use two homophilous graphs Cora and Citeseer and two heterophilic graphs Deezer-Europe and Actor. These graph datasets are all public available at Pytorch Geometric. The Deezer dataset is provided from Non-Homophilous Benchmark, and the Actor (also called Film) dataset is provided by Geom-GCN.

• Large Graph Datasets: we use OGBN-Proteins and Amazon2M as two large-scale datasets. These datasets are available at OGB. The original Amazon2M is collected by ClusterGCN and later used to construct the OGBN-Products.

• Graph-Enhanced Classification: we also consider two datasets without input graphs, i.e., Mini-Imagenet and 20News-Group for image and text classification, respectively. The Mini-Imagenet dataset is provided by Matching Network, and 20News-Group is available at Scikit-Learn

We also provide an easy access to common datasets including what we used (except two large ones) in the Google drive:

https://drive.google.com/drive/folders/1sWIlpeT_TaZstNB5MWrXgLmh522kx4XV?usp=share_link

The two large datasets can be automatically downloaded by our dataloader functions in dataset.py

1. Install the required package according to requirements.txt

2. Download the datasets into a folder ../data

3. To run the training and evaluation on eight datasets we used, one can use the scripts in run.sh:

# node classification on small datasets
python main.py --dataset cora --rand_split --metric acc --method nodeformer --lr 0.001 \
--weight_decay 5e-3 --num_layers 2 --hidden_channels 32 --num_heads 4 --rb_order 2 --rb_trans sigmoid \
--lamda 1.0 --M 30 --K 10 --use_bn --use_residual --use_gumbel --runs 5 --epochs 1000 --device 0

# node classification on large datasets
python main-batch.py --dataset ogbn-proteins --metric rocauc --method nodeformer --lr 1e-2 \
--weight_decay 0. --num_layers 3 --hidden_channels 64 --num_heads 1 --rb_order 1 --rb_trans identity \
--lamda 0.1 --M 50 --K 5 --use_bn --use_residual --use_gumbel --use_act --use_jk --batch_size 10000 \
--runs 5 --epochs 1000 --eval_step 9 --device 0

# image and text datasets
python main.py --dataset mini --metric acc --rand_split --method nodeformer --lr 0.001\
--weight_decay 5e-3 --num_layers 2 --hidden_channels 128 --num_heads 6\
--rb_order 2 --rb_trans sigmoid --lamda 1.0 --M 30 --K 10 --use_bn --use_residual --use_gumbel \
--run 5 --epochs 300 --device 0

• Release the codes and model implementation.

• Provide a detailed tutorial for NodeFormer.

• Provide an example demo for emb/structure visualization.

Our work takes an initial step for exploring how to build a scalable graph Transformer model for node classification, and we also believe there exists ample room for further research and development as future works. One can also use our implementation kernelized_softmax() and kernelized_gumbel_softmax() for related projects concerning e.g., structure learning and communication, where the scalability matters.

If you find our codes useful, please consider citing our work

      @inproceedings{wu2022nodeformer,
      title = {NodeFormer: A Scalable Graph Structure Learning Transformer for Node Classification},
      author = {Qitian Wu and Wentao Zhao and Zenan Li and David Wipf and Junchi Yan},
      booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
      year = {2022}
      }

We refer to the Performer work for the implementation of softmax kernel computation, and the pipeline for training and preprocessing is developed on basis of the Non-Homophilous Benchmark project.