Attributed Graph Clustering via Graph Convolution Method with Normal Signless Laplacian Filter

Attributed graphs are widely used in real networks, such as social networks, economic networks, and information networks (Web or citations). In recent years, some algorithms based on graph convolutional networks have achieved good results in attributed graph clustering, because they combine structural information with node attributes. However, these algorithms still have some shortcomings in two aspects. First, in smoothing node features, the filtering effect of the graph convolution filter is insufficient; Second, the number of convolution layers is difficult to determine. If it is too small, the global structure may not be captured, and it tends to fall into a local optimum; If it is too large, it will cause over-smoothing and increase additional time and space costs. A graph convolution method with a Normal Signless Laplacian Filter (NSLF) for attributed graph clustering is proposed in this paper. Firstly, a new graph filtering technique based on the Normal Signless Laplacian filter is introduced to achieve smooth node representation and an adaptive selection strategy is developed to determine the number of convolution layers, which combines intra-cluster tightness with inter-cluster separation. Finally, we conducted experiments on five public benchmark datasets to verify the effect of the graph convolution method with a normal signless Laplacian filter on node clustering tasks. The experimental results show that our proposed algorithm not only superior to other state-of-the-art algorithms in terms of several key indicators (compared with the latest method DCP-DEC, it averages more than 7.2% on the dataset Cora and more than 6.2% on PubMed, and also fails to get clustering results on large graphs due to too many neural network parameters that DCP-DEC needs to train). It also has a significant advantage over other deep learning methods in terms of time, as it does not require training neural network parameters.