Representation Then Augmentation: Wide Graph Clustering Network with Multi-Order Filter Fusion and Double-Level Contrastive Learning

Youqing Wang, Tianxiang Zhao, Mingliang Cui, Junbin Gao, Li Liang, Jipeng Guo

Abstract: Deep graph contrastive clustering has attracted widespread attentions due to its self-supervised representation learning paradigm and superior clustering performance. Although, two challenges emerge and result in high computational costs. Most existing contrastive methods adopt the data augmentation and then representation learning strategy, where representation learning with trainable graph convolution is coupled with complex and fixed data augmentation, inevitably limiting the efficiency and flexibility. The similarity metric between positive-negative sample pairs is complex and contrastive objective is partial, limiting the discriminability of representation learning. To solve these challenges, a novel wide graph clustering network (WGCN) adhering to representation and then augmentation framework is proposed, which mainly consists of multi-order filter fusion (MFF) and double-level contrastive learning (DCL) modules. Specifically, the MFF module integrates multi-order low-pass filters to extract smooth and multi-scale topological features, utilizing self-attention fusion to reduce redundancy and obtain comprehensive embedding representation. Further, the DCL module constructs two augmented views by the parallel parameter-unshared Siamese encoders rather than complex augmentations on graph. To achieve simple yet effective self-supervised learning, representation self-supervision and structural consistency oriented double-level contrastive loss is designed, where representation self-supervision maximizes the agreement between pairwise augmented embedding representations and structural consistency promotes the mutual information correlation between appending neighborhoods with similar semantics. Extensive experiments on six benchmark datasets demonstrate the superiority of the proposed WGCN, especially highlighting its time-saving characteristic. The code could be available in the https://github.com/TianxiangZhao0474/WGCN.

External IDs:doi:10.1109/jas.2025.125564