Structural Entropy Guided Graph Hierarchical Pooling

Authors: Junran Wu, Xueyuan Chen, Ke Xu, Shangzhe Li

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The results show that SEP outperforms state-of-the-art graph pooling methods on graph classification benchmarks and obtains superior performance on node classifications.
Researcher Affiliation Academia 1State Key Lab of Software Development Environment, Beihang University, Beijing, 100191, China 2School of Mathematical Science, Beihang University, Beijing, 100191, China. Correspondence to: Shangzhe Li <shangzheli@buaa.edu.cn>.
Pseudocode Yes Algorithm 1 Coding tree with height k via structural entropy minimization Input: a graph G = (V, E), a positive integer k > 1 Output: a coding tree T with height k
Open Source Code Yes The implementation of Algorithm 1, SEP-G and SEP-N can be found at https://github.com/Wu-Junran/SEP.
Open Datasets Yes Seven benchmarks for graph classification are selected from TU datasets (Morris et al., 2020). Specifically, we employ three social network datasets, including IMDB-BINARY, IMDB-MULTI, and COLLAB; and four bioinformatics datasets, including MUTAG, PROTEINS, D&D and NCI1. We evaluate SEP-N under the transductive learning setting, which includes three datasets Cora, Citeseer and Pubmed (Sen et al., 2008).
Dataset Splits Yes Following (Xu et al., 2019; Lee et al., 2019), 10-fold cross-validation is conducted, and we present the average accuracies achieved to validate the performance of SEP-G in graph classification. In addition, we use the 10 percent of the training data as a validation data following the fair comparison setup (Errica et al., 2020). The designated training/validation/testing splits on Cora, Citeseer and Pubmed are adopted, that is, training set has 20 nodes for each class, validation set has 500 nodes and testing set has 1,000 nodes.
Hardware Specification No The paper does not provide specific hardware details used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers needed to replicate the experiment.
Experiment Setup Yes For model configuration, the pooling ratio of all models is set to 25%, the learning rate is set to 5 × 10^-3, and hidden size is set to 32. For model configuration, the learning rate is set to 5 × 10^-4, the hidden size is set {64, 128}, the batch size is set {32, 128}, weight decay is set to 1 × 10^-4, and dropout rate is set {0, 0.5}. For model configuration, the learning rate is set to 0.01, the hidden size is set {16, 32, 128, 256}, weight decay is set {0.02, 5 × 10^-4}, and dropout rate for each layer is set {0, . . . , 0.9}.