Research On Attributed Network Representation Learning

With the deep development of Web3.0,large number of social networks are accumulated.How to represent the network becomes a fundamental research of social analysis tasks.Thus,network representation learning has attracted much attention in recent years.Network embedding aims at learning a low-dimensional dense vector for each node in the network.Usually,it starts from properties or characteristics of network and learn an embedding as the new feature for each node.Then,these embeddings are taken into various network analysis tasks like node classification or link prediction.A large number of researches are devoted to network representation learning,but several problems exist in current research.On one hand,most existing studies mainly focus on modeling graph structure and neglect the attribute information which is also an effective context information especially for sparse network.On the other hand,though attributed network embedding methods take node attribute into account,the informative relations between nodes and their attributes are still underexploited.Moreover,all these methods can only deal with the homogeneous network.In this paper,we propose a novel framework to employ the abundant relation information for attributed network embedding.To this end,we first explore the relations in homogeneous and heterogeneous networks,and then construct relevant relations between nodes and their attributes from homogeneous and heterogeneous networks.Relations include “nodes sharing same attribute”,“nodes' neighbors' attribute” and so on.With these relations,we learn the network embedding by constraining them.Following the instruction of this idea,we develop two models from two frameworks.For “embedding-constraint” framework,we firstly embed each node to a latent space,and use such relations to constrain them to learn the final representation.For “feature extraction-constraint” framework,we propose a relational graph convolution network(RGCN)method to encode relevant relations in both types of networks and train node embeddings by reconstructing the network.By fully developing these relations,we make the best of the structure and attribute information and learn more effective ndoe representation.We conduct extensive experiments on real world datasets.Results demonstrate the effectiveness of our model on various social network analysis and recommendation tasks.By analysing the merits and drawbacks of the two models,we also propose a direction of future improvement.