Variational Auto-Encoder Based Attributed Network Representation Learning And Deep Embedded Clustering

Information networks are widely used in the real world to describe complex relationships among entities,such as social networks,communication networks,and the World Wide Web.Mining valuable knowledge from networks has become a hot topic in both academia and industry in recent years,and plays a crucial role in a variety of emerging applications across various disciplines.Meanwhile,as more and more data information becomes available,nodes in real world are often associated with attribute information,which may play an important role in many applications.Attributed network representation learning aims to learn low-dimensional latent representations of nodes which can well capture the topology and attribute information of networks at the same time.Most existing methods focus on embedding network structure or node attributes to latent space with the proximity relationship preserved.These methods fail to exploit the dependencies between the latent representation of nodes and observed information,including node attributes and network structure.Moreover,node attributes and network structure are not necessarily positive correlated,i.e.there are edges between two nodes but the attributes may be dissimilar,and vice versa.Most of existing methods do not consider this partial correlation.In this thesis,an unsupervised generation model is proposed to model the independent generation process of the network structure and node attributes.Given latent variables,the node attributes and network structure are conditionally independent,reflecting the partial correlation between two kinds of information.The problem of network representation learning is transformed into a probabilistic inference problem of hidden variable.Based on this,a neural network model based on variational auto-encoder is designed to learn latent representation of each node.Based on above,we find that many nodes in the network will exhibit a macro pattern of class structure.In the same class,nodes tend to connect densely or share common attributes.These patterns are expected to improve network representation learning.Most existing community preserving methods only consider the network structure,but ignore attribute information of nodes.Based on this,this thesis proposes a unified framework that jointly solves node representation learning,class label assignment and class representation learning three tasks together to achieve the gain in performance simultaneously.Finally,we evaluate our model on node classification,link prediction and node clustering using three real-world datasets.The experiment results show that the proposed model can achieve significant gains compared to the baseline algorithms.