Heterogeneous Network Embedding Method Based on Hyperbolic Space

The core of heterogeneous network embedding is to transform nodes and edges into low and dense vector representations while preserving the structural features and semantic information of the original network. The existing embedding methods based on European space can not store the hierarchical structure of the network, nor can capture the rich semantic information contained in the heterogeneous network. To improve the embedding quality of heterogeneous networks and reduce information loss, a hyperbolic space-based embedding method for heterogeneous networks is proposed. Firstly, a random walk guided by meta-path is used to obtain the node neighborhood information and semantic relation of the network. Then the heterogeneous network is embedded into hyperbolic space based on Poincare sphere model, and the Riemann stochastic gradient descent method is used to optimize the model. Experiments on multiple datasets show that the proposed method has better performance than other baseline methods, and preserves more complete network information with fewer dimensions of storage space.