Research On Heterogeneous Network Embedding Based On Random Walk

Heterogeneous network embedding is a low-dimensional dense embedding process of learning network,which is composed of various entities and relations.The embedding solves the bottleneck of high-dimensional sparsity and poor scalability in large-scale network.It becomes a research hotspot.Most of the existing research is based on the meta-path given by the experts to realize random walk.However,with the increase of the nodes,the choice of meta-path is difficult and inflexible.This paper mainly studies the random walk strategy.Considered the structure and semantic of the network,a more flexible random walk strategy is proposed.And it is combined with dynamic heterogeneous network to realize the embedding of dynamic network.The researches are as follows.Firstly,a random walk strategy based on Type & Inner is proposed.Because meta-path requires prior knowledge of experts or predefined meta-path and inflexible.JUST only considers the number of node types and ignores the relationship between node types and topology structure in the real network.Secondly,to realize non-meta-path random walk,a transition probability model is constructed based on Type & Inner strategy.It adjusts the node type and node by parameters.To verify the rationality of the model,property analysis and example are given.To obtain embedding,HNE-RWTI(Heterogeneous Network Embedding Based on Random Walk of Type & Inner)was proposed combining the strategy with Skip-Gram.Thirdly,to solve the influence of dynamic characteristics on embedding and improve its accuracy,DHNE-RWTI(Dynamic Heterogeneous Network Embedding Based on Random Walk of Type & Inner)was proposed by combining HNE-RWTI with LSTM.Finally,the classification and clustering tasks were verified by experiments on DBLP and AMiner datasets.F1-score,Accuracy and modularity were used to verify the correctness of the two models proposed in this paper by comparing them with the classical models and analyzing the sensitivity of parameters.