Research On Dynamic Representation Learning Of Heterogeneous Network

In the real world,many data are graph-structured,and networks have become a common language for modeling and describing these complex patterns,such as social networks,information networks,and biological networks.Analysis and research on these data can obtain potentially valuable information from the large amount of data.Traditional graph representation learning methods are mostly based on adjacency matrices or adjacency lists to represent graph data,which are not suitable for growing and sparse large-scale networks.With the development of deep learning,the network representation learning(NRL)technology has emerged as the times require and has become a research hotspot in the field of data mining.The real networks in life are usually heterogeneous networks,which contain different types of nodes and connections,and exhibit dynamic changing laws,which pose challenges to traditional network representation learning methods.Existing researches on representation learning in heterogeneous networks usually model different semantic information separately,ignoring the knowledge exchange and information fusion between semantics.Moreover,most models for heterogeneous networks are aimed at static networks,which cannot capture the dynamic evolution characteristics of heterogeneous networks,and ignore the evolution rules in the dynamic change process of heterogeneous networks.This thesis constructs a dynamic representation learning model DYHNE for heterogeneous information networks.In terms of heterogeneity modeling,the model divides heterogeneous information networks based on meta-paths and learns the structural and semantic knowledge of different views.Transformer is designed to obtain the correlation between different semantic information,maximizes the fusion and consensus of meta-path information,and encodes the global shared semantics.In terms of dynamicity modeling,this thesis proposes a long-term and short-term dynamic embedding method to solve the coarse-grained problem in network snapshots.The short-term embedding method uses a time decay function,and learns the interaction between nodes based on Newton’s cooling law to ensure that the influence of historical neighbor nodes on the current target node satisfies the trend of time decay.The longterm embedding method adopts the Temporal Convolutional Network(TCN)to retain the effective long-term memory in the network historical features,overcoming the inability of parallel processing by the recurrent neural network in the traditional time series modeling method,and further improves the computational efficiency of the model.DYHNE constructs a temporal graph attention network based on the short-term embedding method,which captures the short-term changes of the network,and constructs a temporal encoder based on the longterm embedding method,which learns the long-term variation law of the network.In the end,we use experiments on three real-world datasets with different time-varying frequencies to verify the effectiveness of our model.The experimental results show that compared with the mainstream network representation learning methods,DYHNE achieves better results in dynamic link prediction and node classification tasks.