TDR-HGN:Residual-enhanced heterogeneous graph networks for topology-driven feature completion

Zhaowei Liu, Yihao Jiang, Yongchao Song, Yao Shan, Wang Kerry, Dong Yang

Keywords: Heterogeneous graph networks, feature completion, topological features, residual networks, meta-path

TL;DR: TDR-HGN is a heterogeneous graph neural network that uses topology-driven feature completion and feature enhancement, combined with meta-path and residual network, and tested on multiple datasets to verify its effectiveness

Abstract: Heterogeneous graphs are composed of multiple types of edges and nodes. The existing heterogeneous graph neural network can be understood as a node feature smoothing process guided by the graph structure, which can accurately simulate complex relationships in the real world. However, due to real-world privacy and data scarcity, some node features are inevitably missing. Furthermore, as model depth increases and multiple types of meta-paths are aggregated, node embeddings tend to be consistent, leading to semantic confusion and overfitting problems. To improve the quality of node embeddings, we propose topology-driven residual boosting network (TDR-HGN). It introduces one-hot encoding and node type encoding to generate initial features, uses topological structure features to guide feature completion, combines residual networks to deal with semantic confusion and over-fitting problems, and builds neighbor-based high-order graph networks through meta-paths to achieve feature enhancement. We conduct extensive experiments on three heterogeneous graph datasets, and the results show that TDR-HGN can significantly improve the performance compared to other methods.

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.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

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: 10400