Reproduce “Structural Deep Clustering Network” on OpenResearch

About this paper

Deyu Bo, Xiao Wang, Chuan Shi, Meiqi Zhu, Emiao Lu, Peng Cui

An end-to-end framework, the Structural Deep Clustering Network (SDCN), integrates intrinsic data features with structural relationships for deep clustering. It achieves superior clustering performance across diverse datasets, showing average improvements of 6% ACC, 17% NMI, and 28% ARI over baselines, by effectively combining autoencoder and Graph Convolutional Network (GCN) modules through a novel delivery operator and dual self-supervision.

Key insights

• The novel "delivery operator" enables the GCN to act as an approximate second-order graph regularization, enriching GCN layers with multi-level intrinsic features from the autoencoder.
• Fusing intrinsic feature representations from an autoencoder with structural information within the GCN module effectively mitigates the GCN over-smoothing problem, allowing for deeper GCNs.
• A dual self-supervised learning strategy, using a shared target distribution to guide both the autoencoder and GCN, results in mutually enhanced and more robust cluster assignments.

Summary by alphaXiv.