Multilayer Graph Contrastive Clustering Network

Abstract

Multilayer graph has garnered plenty of research attention in many areas due to their high utility in modeling interdependent systems. However, clustering of multilayer graph, which aims at dividing the graph nodes into categories or communities, is still at a nascent stage. Existing methods are often limited to exploiting the multiview attributes or multiple networks and ignoring more complex and richer network frameworks. To this end, we propose a generic and effective autoencoder framework for multilayer graph clustering named Multilayer Graph Contrastive Clustering Network (MGCCN). MGCCN consists of three modules: (1)Attention mechanism is applied to better capture the relevance between nodes and neighbors for better node embeddings. (2)To better explore the consistent information in different networks, a contrastive fusion strategy is introduced. (3)MGCCN employs a self-supervised component that iteratively strengthens the node embedding and clustering. Extensive experiments on different types of real-world graph data indicate that our proposed method outperforms stateof-the-art techniques.

Simple Contrastive Graph Clustering

“…Deep graph clustering is a fundamental yet challenging task that aims to reveal the underlying graph structure and divides the nodes into several disjoint groups. According to the learning mechanism, the existing deep graph clustering methods can be roughly categorized into three classes: generative methods [2,4,15,26,27,33,39,40,54], adversarial methods [25,29], and contrastive methods [5,11,20,21,46,47,55]. Our proposed method belongs to the last category.…”

“…Graph learning is becoming increasingly crucial in multimedia applications like facial expression recognition [49], video action recognition [23], and the recommendation system [19] for its good hidden correlation exploiting capability. Among all the directions in graph learning, a fundamental and challenging task, i.e., deep graph clustering, has recently attracted intensive attention [2,5,11,20,21,25,27,33,39,46,47].…”

“…According to the learning mechanism, the existing deep graph clustering methods can be roughly categorized into three classes: generative methods [2,4,15,26,27,33,39,40,54], adversarial methods [25,29], and contrastive methods [5,11,20,21,46,47,55]. In early literature, the generative methods and adversarial methods improve clustering performance by learning cluster-oriented node representations and designing fake sample generation-recognition mechanisms, respectively.…”

“…More recently, thanks to the development of contrastive learning, more consistent and discriminative contrastive loss functions are designed to replace the clustering guided loss function for network training. As a result, the manual trial-and-error problem is alleviated, and the clustering performance is improved [5,11,20,21,46,47,55]. However, complicated data augmentations and time-consuming graph convolutional operation undermine the efficiency of these methods, making them computational time and space consuming (See Section 4.4).…”

“…• We design a novel neighbor-oriented contrastive loss to keep the structural consistency even across views, thus improving the discriminative capability of our network. • Extensive experimental results on seven benchmark datasets demonstrate the superiority and efficiency of the proposed Method Data Augmentation Network Architecture Objective Function Form AGE [5] No Data Augmentation Graph Filters+MLP Cross-Entropy Loss MVGRL [11] Graph Diffusion Un-shared GCNs+ shared MLP InfoMax Loss MCGC [47] No Data Augmentation Graph Filters InfoNCE Loss DCRN [21] Attribute Corruption+Graph Diffusion Auto-Encoder+GCN MSE Loss to Identity Matrix SCAGC [46] Attribute Corruption+Edge Perturbation Shared GCN InfoNCE Loss MGCCN [20] No Data Augmentation Shared GCN InfoNCE Loss SCGC (Ours) Un-shared MLPs+Embedding Corruption Graph Filters+Un-shared MLPs MSE Loss to Adjacent Matrix…”