A Research On Heterogeneous Information Network Representation Algorithm Based On Recurrent Neural Network

Effectively analyzing and mining large-scale Heterogeneous Information Networks(HINs)by adopting network representation learning approaches have received increasing attention.By taking the type information of nodes and edges into consideration,HINs contain abundant semantic and structural information.Such information not only facilities network analysis and downstream tasks to a great extent,but also poses special challenges to well capture such rich information.Traditional representation learning methods on HINs generate node sequences via meta-path guided random walk,and utilize neural language model to learn representation.Though these methods can achieve excellent performance,such methods still utilize the very traditional Skip-Gram model,and ignore the natural features of node sequences.Considering the unique advantage brought by Bidirectional Recurrent Neural Network(Bi-RNN)on the processing of sequence,this paper utilizes Bi-RNN as basic model to study the problem of representation learning for HINs,and proposes a representation learning model for HINs based on Bi-RNN,called RL4 HIN.This paper,first,discusses the potential dependence existed in indirect neighbors of information network and captures such latent dependence via skip-dependence strategy,then studies the different abilities of forward layer and backward layer in Bi-RNN to remain semantic of HINs and balances such difference through a weighted loss function,finally the information transfer and aggregation are carried out by using the idea of graph convolution neural network considering the natural graph structure of information network.Experiments and model analysis,including multi-label node classification,network visualization,scalability analysis and parameter sensitivity analysis,based on two large-scale real-world academic network datasets demonstrate that RL4 HIN significantly outperforms several state-of-the-art network representation learning approaches,and has good adaptability when facing large-scale information network.