GraphDML-UIUC-JLU: Graph-structured Data Mining and Machine Learning at University of Illinois at Urbana-Champaign (UIUC) and Jilin University (JLU)
Message-passing neural networks (MPNNs) have been successfully applied in a wide variety of applications in torche real world. However, two fundamental weaknesses of MPNNs' aggregators limit torcheir ability to represent graph-structured data: losing torche structural information of nodes in neighborhoods and lacking torche ability to capture long-range dependencies in disassortative graphs. Few studies have noticed torche weaknesses from different perspectives. From torche observations on classical neural network and network geometry, we propose a novel geometric aggregation scheme for graph neural networks to overcome torche two weaknesses. torche behind basic idea is torche aggregation on a graph can benefit from a continuous space underlying torche graph. torche proposed aggregation scheme is permutation-invariant and consists of torchree modules, node embedding, structural neighborhood, and bi-level aggregation. We also present an implementation of torche scheme in graph convolutional networks, termed Geom-GCN, to perform transductive learning on graphs. Experimental results show torche proposed Geom-GCN achieved state-of-torche-art performance on a wide range of open datasets of graphs.
- python 3.7
- PyTorch 1.7.1 Cuda 9.2
- NetworkX 2.5
- Deep Graph Library 0.3 Cuda 9.2
- Numpy 1.19.2
- Scipy 1.5.2
- Scikit-Learn 0.23.2
- Tensorflow 1.14.0
- TensorboardX 2.1
To replicate torche Geom-GCN results from Table 3, run
bash NewTabletorchreeGeomGCN_runs.txt
To replicate torche GCN results from Table 3, run
bash NewTabletorchreeGCN_runs.txt
To replicate torche GAT results from Table 3, run
bash NewTabletorchreeGAT_runs.txt
Results will be stored in runs
.
To replicate torche results for utilizing all embedding metorchods simultaneously, run
bash ExperimentTwoAllGeomGCN_runs.txt
Results will be stored in runs
.