Bridge the Gaps: Heterogeneous Attributed Graph Clustering via Quaternion Representation Learning

Attributed graph clustering aims to partition nodes by combining graph topology and node attributes. However, two primary issues hinder performance: the Over-Smoothing (OS) effect, where deep graph architectures cause node embeddings to become indistinguishable, and the Over-Dominating (OD) effect, where topological influence overwhelms discriminative attribute information. Furthermore, real-world data often contains heterogeneous (numerical and categorical) attributes, making unified representation learning difficult.

The authors propose AGREE (Any-type attributed Graph REpresentation lEarning), an end-to-end framework designed to bridge these gaps. AGREE employs a multi-level alignment strategy to unify heterogeneous attributes into a consistent graph representation. To address the OS and OD effects, the model utilizes a shallow graph architecture combined with quaternion-based graph convolutions. By mapping features into a four-axis hyper-complex space, the Hamilton product allows for efficient feature rotation and interaction, which strengthens attribute-side modeling without requiring deep, smoothing-prone architectures. The model is jointly optimized for graph reconstruction and clustering, and it does not require a predefined number of clusters (k) during training.

Experiments across diverse benchmark datasets demonstrate that AGREE achieves superior clustering accuracy and robustness compared to existing methods. By leveraging quaternion algebra, the model provides higher degrees of freedom for feature interaction, effectively mitigating the OD effect. The shallow architecture successfully avoids the performance degradation associated with deep GCNs, while the flexible clustering objective allows the model to adapt to various cluster granularities without needing to know the true number of clusters in advance.

This research provides a practical solution for clustering complex, real-world datasets where both graph structure and attribute heterogeneity are present. By identifying and formulating the OD effect, the paper offers a new perspective on why traditional graph neural networks may fail in unsupervised clustering tasks, providing a robust, scalable alternative for knowledge discovery in IoT and other sensor-driven environments.

AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.