SAME: Uncovering GNN Black Box with Structure-aware Shapley-based Multipiece Explanations
Authors: Ziyuan Ye, Rihan Huang, Qilin Wu, Quanying Liu
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on real-world and synthetic benchmarks show that SAME improves the previous state-of-the-art fidelity performance by 12.9% on BBBP, 7.01% on MUTAG, 42.3% on Graph-SST2, 38.9% on Graph-SST5, 11.3% on BA-2Motifs and 18.2% on BA-Shapes under the same testing condition. |
| Researcher Affiliation | Academia | Ziyuan Ye1,2,*, Rihan Huang1,3, , Qilin Wu1,4, Quanying Liu1, 1Southern University of Science and Technology, 2The Hong Kong Polytechnic University, 3King Abdullah University of Science and Technology, 4Carnegie Mellon University |
| Pseudocode | Yes | Algorithm 1 outlines the process of using Monte Carlo Tree Search (MCTS) to find a set of important substructures, as mentioned in Section 3.1 of the main manuscript. Algorithm 2 and 3 illustrate the method to discover an optimal explanation from these important substructures, as described in Section 3.2. Algorithm 4 provides a detailed description of how the importance of an explanation is assessed using the structure-aware Shapley value, which was proposed in Section 2.1 of your document. |
| Open Source Code | Yes | Code is available at https://github.com/same2023neurips/same. |
| Open Datasets | Yes | The experiments are conducted on six datasets with diverse categories, including molecular graphs (e.g., BBBP [33] and MUTAG [7]), sentiment graphs (e.g., Graph-SST2 and Graph-SST5 [29]) and synthetic Barabási-Albert graphs (e.g., BA-2Motifs [19] and BA-Shapes [35]). |
| Dataset Splits | No | No specific percentages or absolute counts for training, validation, and test splits are provided for most datasets. The paper mentions 'all the data (train, evaluation and test data) is used for MUTAG dataset' but doesn't specify the split ratios or methods for other datasets. |
| Hardware Specification | Yes | Our experiments are conducted on a single Nvidia V100 GPU with an Intel Xeon Gold 5218 CPU. |
| Software Dependencies | No | All methods are implemented in Py Torch [21] and Py G [10]. Specific version numbers for these software dependencies are not provided. |
| Experiment Setup | Yes | We adhere to the hyperparameter settings for GNN training as described in [13, 15], detailed in Table S2. Table S2: Hyperparameters for training the GCN models on different datasets. All quantitative results are verified under the following conditions. |