Beyond Homophily in Graph Neural Networks: Current Limitations and Effective Designs

Authors: Jiong Zhu, Yujun Yan, Lingxiao Zhao, Mark Heimann, Leman Akoglu, Danai Koutra

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical analysis shows that the identified designs increase the accuracy of GNNs by up to 40% and 27% over models without them on synthetic and real networks with heterophily, respectively, and yield competitive performance under homophily. and Extensive Empirical Evaluation: We empirically analyze our model and competitive existing GNN models on both synthetic and real networks covering the full spectrum of low-to-high homophily
Researcher Affiliation Academia Jiong Zhu University of Michigan jiongzhu@umich.edu Yujun Yan University of Michigan yujunyan@umich.edu Lingxiao Zhao Carnegie Mellon University lingxia1@andrew.cmu.edu Mark Heimann University of Michigan mheimann@umich.edu Leman Akoglu Carnegie Mellon University lakoglu@andrew.cmu.edu Danai Koutra University of Michigan dkoutra@umich.edu
Pseudocode Yes We describe H2GCN, which exemplifies how effectively combining designs D1-D3 can help better adapt to the whole spectrum of low-to-high homophily, while avoiding interference with other designs. It has three stages (Alg. 1, App. D):
Open Source Code Yes We compare it to prior GNN models, and make our code and data available at https://github.com/Gems Lab/H2GCN.
Open Datasets Yes We generate synthetic graphs with various homophily ratios h (Tab. 3) by adopting an approach similar to [16]. and We now evaluate the performance of our model and existing GNNs on a variety of real-world datasets [35, 29, 30, 22, 4, 31] with edge homophily ratio h ranging from strong heterophily to strong homophily, going beyond the traditional Cora, Pubmed and Citeseer graphs that have strong homophily (hence the good performance of existing GNNs on them).
Dataset Splits Yes All methods share the same training, validation and test splits (25%, 25%, 50% per class), and we report the average accuracy and standard deviation (stdev) over three generated graphs per heterophily level and benchmark dataset. For all benchmarks (except Cora-Full), we use the feature vectors, class labels, and 10 random splits (48%/32%/20% of nodes per class for train/validation/test2) provided by [26]. For Cora-Full, we generate 3 random splits, with 25%/25%/50% of nodes per class for train/validation/test.
Hardware Specification Yes We gratefully acknowledge the support of NVIDIA Corporation with the donation of the Quadro P6000 GPU used for this research.
Software Dependencies No The paper does not provide specific software dependency details with version numbers, such as Python or PyTorch versions, or specific library versions used for implementation or experiments.
Experiment Setup Yes We tune all the models on the same train/validation splits (see App. F for details). Appendix F states: 'We perform hyperparameter tuning over the following parameters for all models: learning rate (0.01 for GCN, GAT, GCN-Cheby, MixHop, GraphSAGE, H2GCN; 0.001 for MLP), weight decay (0.0005 for all), dropout (0.5 for all except MLP which is 0.0), and number of hidden units (256 for all). We train all models for a maximum of 1000 epochs, using Adam optimizer (with momentum 0.9 and decay 0.999), and full batch.'