Attribute Graph Deep Clustering Based On Contrastive Self-supervised Learning

Many systems in the real world can be modeled as graphs or networks,where nodes represent data objects and edges represent relationships between data objects.In addition,the nodes themselves often contain rich attribute information.Like a paper citation network,the node attribute is the bag of words associated with the paper.With the development of graph neural networks,how to effectively learn non-European data such as graphs has become one of the main hotspots in deep learning research today.Existing deep learning methods on graphs often focus on node representation learning.The learned representation vector can be used for downstream node classification,clustering or link prediction tasks,but the learned representation vector may not be optimal for a specific downstream,especially for the unsupervised node clustering task.Since self-supervised learning can effectively learn the representation of nodes,and combined with clustering goals can achieve end-to-end node clustering tasks,how to introduce self-supervised learning into graph depth clustering and give effective map depth clustering methods has become an emerging research topic.For the task of deep clustering of nodes on the graph,the main work of this paper is as follows:(1)Combining the latest contrastive learning method with the node clustering task,the effectiveness of the graph depth clustering methods GRACE-DEC and MVGRL-DEC based on contrastive learning is realized and verified.A large number of comparative experiments are carried out on four real network data,and the experimental results show that the deep clustering model based on contrastive self-supervision proposed in this paper is significantly better than the generative self-supervised deep clustering method.(2)In view of the problem that the existing graph deep clustering model does not consider the detection of overlapping nodes,this paper integrates fuzzy clustering into the deep clustering model based on contrastive self-supervision,and proposes a deep fuzzy clustering method GRACE-FDEC.The validity of the deep clustering model of overlapping nodes in the graph is verified on multiple artificial and real data.Experiments are carried out on multiple overlapping network data,and the results show that the endto-end deep fuzzy clustering method that integrates clustering targets is better than the two-step method of learning node representation vectors first and then clustering and the traditional overlapping node detection method.