DEGREE: Decomposition Based Explanation for Graph Neural Networks
Authors: Qizhang Feng, Ninghao Liu, Fan Yang, Ruixiang Tang, Mengnan Du, Xia Hu
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, we conduct quantitative and qualitative experiments on synthetic and real-world datasets to demonstrate the effectiveness of DEGREE on node classification and graph classification tasks. |
| Researcher Affiliation | Academia | 1Department of Computer Science and Engineering, Texas A&M University 2Department of Computer Science, University of Georgia 3Department of Computer Science, Rice University |
| Pseudocode | Yes | We conclude the computation steps of subgraph-level explanation (Sec 4) in Algorithm 1, 2 and 3. |
| Open Source Code | Yes | The code can be found at https://anonymous.4open.science/r/DEGREE-3128. |
| Open Datasets | Yes | Following the setting in previous work (Ying et al., 2019), we adopt both synthetic datasets and real-world datasets. The statistic of all datasets are given in Sec A in the Appendix. BA-Shapes. BA-Shapes is a unitary graph based on a 300-node Barab asi-Albert (BA) graph (Barab asi & Albert, 1999). MUTAG. It is a dataset with 4,337 molecule graphs. Every graph is labeled according to their mutagenic effect on the bacterium. As discussed in (Debnath et al., 1991), the molecule with chemical group NH2 or NO2 and carbon rings are known to be mutagenic. Graph-SST2. It is a dataset of 70,042 sentiment graphs, which are converted through Biaffine parser (Liu et al., 2021). The node features are initialized as the pre-trained BERT word embeddings (Devlin et al., 2019). |
| Dataset Splits | Yes | For all datasets, we use a train/validation/test split of 80%/10%/10%. |
| Hardware Specification | Yes | CPU: AMD EPYC 7282 16-Core Processor. GPU: Ge Force RTX 3090 NVIDIA-SMI: 460.32.03 Driver Version: 460.32.03 CUDA Version: 11.2. |
| Software Dependencies | Yes | CUDA Version: 11.2. |
| Experiment Setup | Yes | We set the number of GNN layers to 3 for all datasets, except for the Tree-Grid dataset where it is 4. Since the 3-hop neighbors of some target nodes has only in-motif nodes (no negative samples). For all synthetic datasets, the GCN model is trained for 1,000 epochs and the GAT model is trained for 200 epochs. For MUTAG dataset, the GCN and GAT model is trained 30 epochs. For Graph-SST2 dataset, the GCN model is trained 10 epochs. We use Adam optimizer and set the learning rate to 0.005, the other parameters remain at their default values. |