KDGCN: A Kernel-based Double-level Graph Convolution Network for Semi-supervised Graph Classification with Scarce Labels
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: graph neural network, graph classification, semi-supervised learning
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Graph classification, which is significant in various fields, often faces the challenge of label scarcity. Under such a scenario, supervised methods based on graph neural networks do not perform well because they only utilize information from labeled data. Meanwhile, semi-supervised methods based on graph contrastive learning often yield complex models as well as elaborate hyperparameter-tuning. In this work, we present a novel semi-supervised graph classification method, which combines GCN modules with graph kernels such as Weisfeiler-Lehman subtree kernel. First, we use a GCN module as well as a readout operation to attain a graph feature vector for each graph in the dataset. Then, we view the graphs as meta-nodes of a supergraph constructed by a graph kernel among graphs. Finally, we use another GCN module, whose inputs are the graph feature vectors, to learn meta-node representations over the supergraph in a semi-supervised manner. Note that the two GCN modules are optimized jointly. Compared to contrastive learning based semi-supervised graph classification methods, our method has fewer hyperparameters and is easier to implement. Experiments on seven benchmark datasets demonstrate the effectiveness of our method in comparison to many baselines including supervised GCNs, label propagation, graph contrastive learning, etc.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2196
Loading