Graph Convolutional Neural Networks Based On Graph Disentanglement Learning

Graph structure data,which commonly exists in the real world,describes a group of entities and the relationships between these entities as nodes and edges respectively.As a general network model for graph structure data learning,graph neural networks,including graph convolutional neural networks,focus on the propagation and aggregation of information of nodes according to graph topology.The networks have been successfully applied in many fields,such as recommendation system,social network,biomolecular field,and etc.However,when the depth of the graph convolutional neural network increases,the performance decreases rapidly with high degree of similarities between nodes,that is,over-smoothing.The phenomenon of over-smoothing generally exists in graph convolutional neural networks,which is attributed to its message passing mechanism that all neighborhood nodes of a target node are taken as a whole and all neighborhood information is aggregated in a unified way to update the target node representations.The aggregation step ignores the subtle differences between nodes,which usually contain a variety of complex and implicit interactions.For example,a person in a social network is associated with others due to a variety of latent factors,including work,school,relatives,and etc.,which can not be recognized by the existing graph convolutional neural networks.Disentangled graph neural networks provide a novel perspective for over-smoothing and introduce graph disentangled representation learning method.The method decomposes node neighborhood into several identified latent factors and learns graph data so that enhancing the diversity of node representations and easing over-smoothing.Recently,there are many challenges for recent graph disentanglement learning methods:(1)When consistency of inner latent factor disappears,it will weaken the inner correlation of the latent factor and the interpretability of node representations.(2)When independence of potential factors is insuffieiect,it will lead to redundant dependences between latent factors and damage diversity of node representations.In this paper,based on the above problems,we propose a disentangled graph convolutional neural network based on latent bottleneck and a global disentangled graph convolutional neural network based on graph topological metric.The main contributions of this paper are as follows:(1)We propose a graph convolutional neural network based on latent bottleneck is proposed for node-level graph disentangled representation learning.This model introduces information bottleneck and identifies latent bottleneck from the input node information for latent node representation generation with restriction of the transmission of latent graph information according to specific latent factor.Meanwhile,these identified latent bottlenecks promote independence between latent factors.Through the definition of latent bottleneck distributions,the inner consistency of specific latent factor between different nodes is guaranteed.(2)There are some defects on the disentangled graph convolutional neural network based on the latent bottleneck.In this paper,we adopt a generative model to eatablish the correspondence between input node information and latent node representations in order to avoid the loss of partial latent infomation corresponding to specific latent factor.We perform several experiments to evaluate our model on disentanglement performance and prediction performance,including semi-supervised node classification tasks,clustering coefficient analysis,and disentanglement performance analysis.The experimental results prove that graph disentanglement learning can enhance node representation diversity and then relieve over-smoothing.(3)We propose a global disentangled graph convolutional neural network based on graph topological metric for graph-level disentangled representation learning.The network decomposes the input graph into several factor graphs respectively corresponding to latent factors.In order to promote disentanglement through graph information flow in each graph convolutional network layer,we introduce Jensen Shannon MI evaluator.Meanwhile,we propose graph topological metric to measure similarities between topologies of factor graphs.The experimental results demonstrate abilitily as a general graph nerual model framework and enhance node representation diversity,effectively.