Scalable Graph Neural Networks via Bidirectional Propagation

Authors: Ming Chen, Zhewei Wei, Bolin Ding, Yaliang Li, Ye Yuan, Xiaoyong Du, Ji-Rong Wen

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental An extensive empirical study demonstrates that GBP achieves state-of-the-art performance with significantly less training/testing time. Most notably, GBP can deliver superior performance on a graph with over 60 million nodes and 1.8 billion edges in less than half an hour on a single machine.
Researcher Affiliation Collaboration Ming Chen Renmin University of China chennnming@ruc.edu.cn Zhewei Wei Renmin University of China zhewei@ruc.edu.cn Bolin Ding Alibaba Group bolin.ding@alibaba-inc.com Yaliang Li Alibaba Group yaliang.li@alibaba-inc.com Ye Yuan Beijing Institute of Technology yuan-ye@bit.edu.cn Xiaoyong Du Renmin University of China duyong@ruc.edu.cn Ji-Rong Wen Renmin University of China jrwen@ruc.edu.cn
Pseudocode Yes Algorithm 1: Bidirectional Propagation Algorithm Input: Graph G, level L, training set Vt, weight coefficients wℓ, convolutional coefficient r, threshold rmax, number of walks per node nr, feature matrix Xn F Output: Embedding matrix Pn F 1 S(ℓ) Sparse (0n n) for ℓ= 0, . . . , L; 2 for each node s Vt do 3 Generate nr random walks from s, each of length L; 4 if The j-th random walk visits node u at the ℓ-th step, ℓ= 0, . . . , L, j = 1, . . . , nr then 5 S(ℓ)(s, u) += 1 nr ; 6 Q(ℓ), R(ℓ) Sparse (0n F ) for ℓ= 1, . . . , L; 7 Q(0) 0n F and R(0) Column Normalized (D r X); 8 for ℓfrom 0 to L 1 do 9 for each u V and k {0, . . . , F 1} with R(ℓ)(u, k) > rmax do 10 for each v N(u) do 11 R(ℓ+1)(v, k) += R(ℓ)(u,k) 12 Q(ℓ)(u, k) R(ℓ)(u, k) and R(ℓ)(u, k) 0; 13 Q(L) R(L); 14 ˆP PL ℓ=0 wℓ Dr Q(ℓ) + Pℓ t=0 S(ℓ t)R(t) ; 15 return Embedding matrix ˆPn F ;
Open Source Code No The paper does not explicitly state that the source code for the described methodology (GBP) is made publicly available, nor does it provide a direct link to a repository for its implementation. It only mentions using the officially released code of baselines.
Open Datasets Yes We use seven open graph datasets with different size: three citation networks Cora, Citeser and Pubmed [25], a Protein-Protein interaction network PPI [11], a customer interaction network Yelp [37], a co-purchasing networks Amazon [8] and a large social network Friendster [34]. Table 2 summarizes the statistics of the datasets.
Dataset Splits Yes Following [15], we apply the standard fixed training/validation/testing split with 20 nodes per class for training, 500 nodes for validation and 1,000 nodes for testing. For each method, we set the number of hidden layers to 2 and take the mean accuracy with the standard deviation after ten runs. ... We use fixed-partition splits for each dataset, following [37, 8] (see the supplementary materials for further details). ... Among the labeled nodes, we use 50,000 nodes (100 from each class) for training, 15,982 for validation, and 25,000 for testing.
Hardware Specification Yes All the experiments are conducted on a machine with an NVIDIA TITAN V GPU (12GB memory), Intel Xeon CPU (2.20 GHz), and 256GB of RAM.
Software Dependencies No The paper states, 'We implement GBP in PyTorch and C++,' but does not provide specific version numbers for PyTorch, C++, or any other software dependencies.
Experiment Setup Yes We set L = 4 across all datasets. Table 3 summaries other hyper-parameters of GBP on different datasets. We use Adam optimizer to train our model, with a maximum of 1000 epochs and a learning rate of 0.01. For a fair comparison, we use the officially released code of each baseline (see the supplementary materials for URL and commit numbers) and perform a grid search to tune hyperparameters for models.