Score-based Explainability for Graph Representations
Authors: Ehsan Hajiramezanali, Sepideh Maleki, Max W Shen, Kangway V. Chuang, Tommaso Biancalani, Gabriele Scalia
TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive qualitative and quantitative experiments demonstrate gr XAI s strong ability to identify subgraphs that effectively explain learned graph representations across various unsupervised tasks and learning algorithms. 5 Experiments: To evaluate the effectiveness of the proposed method, we performed experiments on various datasets, graph learning models, and explainability methods. We evaluated our gr XAI framework on seven datasets in both unsupervised and supervised settings, covering synthetic, biological, citation network, and text data. |
| Researcher Affiliation | Industry | Ehsan Hajiramezanali EMAIL Genentech; Sepideh Maleki EMAIL Genentech; Max W Shen EMAIL Genentech; Kangway V. Chuang EMAIL Genentech; Tommaso Biancalani EMAIL Genentech; Gabriele Scalia EMAIL Genentech. All authors are affiliated with Genentech and their emails end with @gene.com, indicating an industry affiliation. |
| Pseudocode | Yes | Algorithm 1 in Appendix G includes the computation steps for gr Subgraph X. Appendix G gr Subgraph X algorithm: Algorithm 1 The algorithm of gr Subgraph X to calculate Shapley score. |
| Open Source Code | No | The paper does not provide an explicit statement about releasing their code or a specific repository link for the methodology described. It mentions using existing tools and implementations (PyTorch, DIG, Captum) but not their own code for the gr XAI framework. |
| Open Datasets | Yes | To evaluate the effectiveness of the proposed method, we performed experiments on various datasets... MUTAG (Debnath et al., 1991), BBBP (Wu et al., 2018), BACE (Wu et al., 2018), and NCI1 (You et al., 2020) are molecular datasets... BA-Shapes (Yuan et al., 2020; 2021) is a synthetic node classification dataset... The Graph-Twitter (Yuan et al., 2020) dataset is a sentiment graph classification dataset... Cora (Sen et al., 2008) is a citation network for node embedding tasks. |
| Dataset Splits | Yes | For supervised experiments, we used GNN checkpoints from DIG (Liu et al., 2021) as part of Yuan et al. (2020) and did not train any new models... We followed the hyperparameters outlined in the main papers for these methods... As the considered real-world datasets do not provide ground truth for explanations, we follow previous studies (Pope et al., 2019; Xie et al., 2022; Yuan et al., 2020) and adopt Fidelity and Sparsity scores to quantitatively evaluate the explanations. |
| Hardware Specification | Yes | We used Py Torch (Paszke et al., 2019) to develop our gr XAIe framework and conducted experiments on an Nvidia A100 with 80 GB of memory. |
| Software Dependencies | No | The paper mentions several software components like PyTorch (Paszke et al., 2019), DIG (Liu et al., 2021), Subgraph X (Yuan et al., 2021), and Captum (Kokhlikyan et al., 2020). However, it does not specify explicit version numbers for these software packages, only citing their respective publications. |
| Experiment Setup | Yes | We followed the hyperparameters outlined in the main papers for these methods, using 3-layer GINs with an embedding dimension of 32 for Graph CL, 4-layer GINs with an embedding dimension of 512 for Info Graph, and 2-layer GCNs with a node embedding dimension of 128 for GRACE. For gr Subgraph X, we modified the original implementation of Subgraph X (Yuan et al., 2021) and followed the same hyperparameters of the original paper. |