Towards Fine-Grained Explainability for Heterogeneous Graph Neural Network
Authors: Tong Li, Jiale Deng, Yanyan Shen, Luyu Qiu, Huang Yongxiang, Caleb Chen Cao
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical results on various HGNs and heterogeneous graphs show that x Path yields faithful explanations efficiently, outperforming the adaptations of advanced GNN explanation approaches. We conduct experiments using three public heterogeneous graph datasets for node classification tasks. |
| Researcher Affiliation | Collaboration | 1 Shanghai Jiao Tong University 2 Huawei Research Hong Kong |
| Pseudocode | Yes | Algorithm 1: Rewiring algorithm Input: G, P = v, v1, , v L, vt; Algorithm 2: Finding top-K fine-grained explanations Input: vt, MG, maximum iteration Lmax, sample size m, candidate size b |
| Open Source Code | Yes | Source code at https://github.com/LITONG99/x Path. |
| Open Datasets | Yes | We conduct experiments using three public heterogeneous graph datasets for node classification tasks. (1) ACM (Wang et al. 2021a) is an academic network... (2) DBLP (Wang et al. 2021a) is a bibliography graph... (3) IMDB2 is a movie graph with node types: movie (M), director (D), and actor (A). The statistics of the datasets are in Table 1. 2https://www.kaggle.com/carolzhangdc/imdb-5000-moviedataset |
| Dataset Splits | Yes | To train these models, we split each dataset into training/validation/test sets and we randomly reserve 1000/1000 samples for validation/test. |
| Hardware Specification | Yes | We conducted all the experiments on a server equipped with Intel(R) Xeon(R) Silver 4110 CPU, 128GB Memory, and a Nvidia Ge Force RTX 2080 Ti GPU (12GB Memory). |
| Software Dependencies | No | The paper states "We implemented x Path with Py Torch" but does not provide specific version numbers for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | For x Path, we tune the hyperparameters (b, m) of ACM to (5, 5), DBLP with SIM2 to (10, 10), and others to (2, 10). The node embedding size is set to 32 in all the models. To train these models, we split each dataset into training/validation/test sets and we randomly reserve 1000/1000 samples for validation/test. We train HGT with 2000 samples and train Simple HGN with 60 samples per label for training efficiency. As suggested by the previous work (Ying et al. 2019), we set the node number to 5 to avoid overwhelming users. |