Learning Network And Network Stream Representations With Group Similarity Constraints

With the popularity and rapid development of the Internet,nowadays,the society has entered the era of information explosion.The information in society is ubiquitous,and the interaction and association of information makes a huge information network,which has diverse forms.The huge network graph contains a lot of potential information that needs to be discovered and exploited.In recent years,the network representation techniques play an important role in the task of network analysis,attracting more and more attention from the academic and industry.The network representation techniques map the network graph structure to the low-dimensional vector space,and learn the vector representation of each node,which can be applied to many network analysis tasks,such as link prediction,community discovery,and anomaly detection.In this paper,we propose a network representation learning algorithm based on the similarity of population.Inspired by the neighborhood of a node in the network graph has some similarity constraint,which called population similarity.Firstly the paper defines the similarity of nodes,then models population similarity,constructs the population similarity loss function,finally applied it to the representation learning of homogeneous and heterogeneous networks.Experiments show that the population similarity constraints have a positive effect on network representation learning.Then we propose a network stream representation learning algorithm for truncated neighbors.According to the large network stream representation learning in the real world.The realistic network graph structure is very large,with a great number of nodes and edges,tens of millions or even hundreds of millions.The memory of a single compute node,even for the disk,is not enough to load such magnitude data,and the calculation time is too long.It keeps a fixed number of neighbors for each node,maintaining the second order similarity of the network structure to a certain extent.Finally,this paper conducts the network analysis task experiments on multiple public data sets,including the node classification of social networks and rating prediction in recommendation scenarios.By comparing with many algorithms,it is proved that theproposed algorithm has better performance and robustness.