Representation Learning Of Dynamic Heterogeneous Social Networks

As the amount of data grows,the scale of the social network continues to expand.The adjacency matrix of the network occupies a large storage space,and cannot fully reflect the overall structural information and semantic information of the network.It cannot be directly applied to traditional machine learning algorithms,so the research of social network's representation learning has important practical significance.The long-term accumulated historical data can construct a dynamic social relationship network,reflecting the dynamic evolution of social network;the fusion of multi-source heterogeneous information can construct a heterogeneous social network,reflecting multiple relationships between various nodes in a social network.The existing research only expresses the representation of one of the dynamic or heterogeneous characteristics of the network.It is still difficult to combine the two aspects of the same vector space to make the embedded result reflect the evolution information and heterogeneous relationship information of the network.We proposes a dynamic heterogeneous social relationship network representation learning algorithm DHIN2Vec.The main work and contributions of this paper are as follows:(1)It is difficult to combine dynamic heterogeneous information and to ensure that the two features are in the same vector space.The combination of dynamic evolution information and heterogeneous information is studied,including serial,parallel and fusion ideas.This paper adopts the fusion idea,uses Meta Path and random walk principle to generate training samples,which reflects the connection between heterogeneous nodes in the network.Refer to Word2Vec,we use neural network activated by the one-hot encoding.Based on LSTM neural network structure,we learn dynamic evolution relationship of the network,a DHIN2Vec model based on deep learning is proposed.DHIN2Vec fits the embedded representation of node pairs in a heterogeneous relationship at different times in a neural network,learning the embedded representation of a dynamic heterogeneous social network.(2)A DBLP academic dynamic heterogeneous social relationship network is constructed for the vacancy of network data that meets both dynamic and heterogeneous characteristics.There are three types of heterogeneous nodes,including authors,articles,and conferences totaling 41430 nodes.Four types of heterogeneous edges are collected in the network,including creation,citation,cooperation,and conferences,totaling 146231 links.These nodes are continuously distributed for 15 years and are divided into five intervals.(3)According to the different semantic problems of heterogeneous edges in the network,experiments have been carried out to verify the edges that do not change with time,such as the creation side between the author and the article,which has been retained in the network at the subsequent moment since the moment of occurrence,allowing nodes to Embedding means having more realistic semantic information,indicating better learning results.In addition,the experiment verified that if the conference node is added to the DBLP network as additional heterogeneous information,DHIN2Vec has better performance.So the effectiveness of heterogeneous information in DHIN2Vec is proved.Compared with typical algorithms such as Deep Walk,Node2Vec,TNE,Dynamic Triadic,and HIN2Vec,DHIN2Vec has a FI-Score of 0.9117 in the link prediction task of the latest time network,which is 0.0534 higher than the best performing Dynamic Triadic in the benchmark algorithm,and has an increase of 6.27%.In node prediction,DHIN2Vec's Macro-F1 is 0.521,which is 0.014 higher than Dynamic Triadic,an increase of 2.76%.The DHIN2Vec fusion embeds both dynamic and heterogeneous data,significantly improving the performance of link prediction and node prediction applications.