HA-SCN: Learning Hierarchical Aligned Subtree Convolutional Networks for Graph Classification

Authors: Xinya Qin, Lu Bai, Lixin Cui, Ming Li, Hangyuan Du, Yue Wang, Edwin Hancock

IJCAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the proposed HA-SCN model on seven benchmark graph classification datasets of two categories: bioinformatics (Bio) and social networks (SN). Detailed descriptions are shown in Table 1. We set the number of nodes in the prototype graph according to the scale of the input graphs. Specifically, we set it to 16 for the MUTAG dataset, 64 for the D&D and REDB datasets, and 32 for the remaining datasets. For each aligned graph, we choose K = 2 and m = 3, constructing a 3-layer binary tree for each node. Our HA-SCN model consists of five layers: one input layer, two subtree convolution pooling layers for extracting node feature information, one MLP layer for integrating graph-level information, and one output layer. In the convolution pooling layers, we set the number of filters to 32. We use the Adam optimization algorithm [Kingma and Ba, 2014] for gradient descent. All weights are randomly initialized from a normal distribution with mean zero and variance 0.01, and we adopt Re LU as the activation function and apply average pooling for readout. To evaluate the classification performance, we perform 10-fold cross-validation, using nine folds for training and one fold for testing. The experiment is repeated 10 times, and we report the average classification accuracies along with the standard errors.
Researcher Affiliation Academia 1School of Artificial Intelligence, Beijing Normal University, Beijing, China 2School of Information, Central University of Finance and Economics, Beijing, China 3Zhejiang Institute of Optoelectronics, Jinhua, China 4Zhejiang Key Laboratory of Intelligent Education Technology and Application, Zhejiang Normal University, Jinhua, China 5School of Computer and Information Technology, Shanxi University, Taiyuan, China 6Department of Computer Science, University of York, York, United Kingdom. Xinya EMAIL, EMAIL, EMAIL.
Pseudocode Yes Algorithm 1 Mapping Graph to Tree Input: Graph, receptive field size m + 1, Page Rank algorithm, graph grafting, graph pruning, depth K Output: K-level m-ary tree for each node 1: Initialize graph structures. 2: Compute Page Rank values for each node. 3: Construct an m-ary tree for each node using graph grafting and pruning algorithms. 4: for i = 2 i K do 5: Replace the leaf nodes of the i-level m-ary tree with their own neighborhood m-ary trees. 6: end for 7: return K-level m-ary tree for each node.
Open Source Code Yes Code is available on Git Hub.1 1https://github.com/Xiaoqin0421/HA-SCN
Open Datasets Yes We evaluate the proposed HA-SCN model on seven benchmark graph classification datasets of two categories: bioinformatics (Bio) and social networks (SN). Detailed descriptions are shown in Table 1.
Dataset Splits Yes To evaluate the classification performance, we perform 10-fold cross-validation, using nine folds for training and one fold for testing. The experiment is repeated 10 times, and we report the average classification accuracies along with the standard errors.
Hardware Specification No No specific hardware details (like GPU/CPU models, memory, or cloud instances) are provided in the paper.
Software Dependencies No The paper mentions the use of the Adam optimization algorithm and ReLU activation function, but does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes We set the number of nodes in the prototype graph according to the scale of the input graphs. Specifically, we set it to 16 for the MUTAG dataset, 64 for the D&D and REDB datasets, and 32 for the remaining datasets. For each aligned graph, we choose K = 2 and m = 3, constructing a 3-layer binary tree for each node. Our HA-SCN model consists of five layers: one input layer, two subtree convolution pooling layers for extracting node feature information, one MLP layer for integrating graph-level information, and one output layer. In the convolution pooling layers, we set the number of filters to 32. We use the Adam optimization algorithm [Kingma and Ba, 2014] for gradient descent. All weights are randomly initialized from a normal distribution with mean zero and variance 0.01, and we adopt Re LU as the activation function and apply average pooling for readout. To evaluate the classification performance, we perform 10-fold cross-validation, using nine folds for training and one fold for testing.