Research On Heterogeneous Social Network Representation Learning Combined With Community Structure Information

In recent years,large-scale social networks represented by Weibo,Facebook and WeChat have developed rapidly,resulting in massive social network data.Massive social network data makes the classic network representation methods encounter bottleneck in data mining of the network.The classic network representation method has problems such as excessive space occupation and inconvenient data form.As a way to solve these problems,network representation learning has received more and more researchers' attention in recent years.The purpose of network representation learning is to map nodes in the network into a low-dimensional vector space,and represent each network node as a feature vector,so that the structural information of the original network can be contained in the vector.Network representation learning makes network data easier to store by making network nodes a low-dimensional vector,and network data in vector form is also easier to use as input to traditional machine learning algorithms.However,the effect of the network representation learning algorithm will quickly decrease as the social network becomes sparse.Because if the social network is sparse,there is no direct or indirect connection among many users in the network,and the network representation learning algorithm cannot determine the distance among the user vectors in the vector space.In order to solve this problem,this paper makes use of the attribute information of user nodes in social networks.By extracting various attributes of the user,new attribute nodes are formed and added to the social network to form a new heterogeneous social network.In heterogeneous social networks,users are linked to each other through different types of attribute nodes and link relationships,which enriches the semantic relationship,thus effectively alleviating the sparseness of social networks.However,the current network representation learning algorithm mainly focuses on the representation learning of the homogeneous network,that is,the nodes and the link relationships in the network are all of the same type.Different heterogeneous social networks contain different types of nodes and links.Different types of links between user nodes contain different semantic relationships.Users may have different distances under different semantic relationships.The network representation learning algorithms designed for homogeneous networks do not necessarily well learn the vector representation of users in the social network.In order to solve this problem,this paper proposes a user structure similarity calculation method based on meta-structure,which uses different types of meta-structures in heterogeneous social networks to more accurately describe the semantic relationship between users in different scenarios.Then use the stacked denoising autoencoder to fuse multiple relationship information to learn the low-dimensional representation of users in heterogeneous social networks.In addition,there are often obvious community structures in social networks.If the community structure of social networks can be learned during the learning process of heterogeneous social networks,the learned network node vectors can reflect the structure of the original network.Features.The obtained network node vector can also exert better effects in subsequent network data mining tasks,such as network node classification,link prediction,and the like.Based on this idea,we propose a network representation learning algorithm that combines community structure information.The algorithm learns the community information of each network node by comparing the community in the social network to the theme in natural language and using the topic model method in natural language processing.And add it to the network to represent the process of learning,and finally learn the final representation of users in the social network.Finally,we demonstrate the effectiveness of the proposed algorithm by experiments on three real data sets.The main contributions and innovations of this paper are as follows:(1)In order to solve the problem that the social network is too sparse to affect the network representation learning effect,this paper reduces the sparseness of social networks by introducing more abundant information.By mining the attribute information of the user node and joining the social network,a new heterogeneous social network is constructed.(2)For the homogeneous network,the learning algorithm is not applicable to heterogeneous social networks,and the problem of rich semantic information in the network cannot be accurately learned.This paper proposes a kind of computing similarity in heterogeneous social networks by using the meta structure in heterogeneous social networks.This method is used to mine different types of semantic information between users in a heterogeneous social network.The deep neural network is then used to fuse multiple semantic information to learn the node representation of the network.(3)In order to learn the community structure information of heterogeneous social networks in network representation learning and get more accurate user representation,this paper compares the community in social network to the theme in natural language,and then draws on the LDA topic model algorithm in natural language processing method.To learn the user's community information and apply it to the network to represent learning.