Dynamic Heterogeneous Network Link Prediction Based on Incremental Node Embedding

In a dynamic heterogeneous network, the number of nodes in the network will continue to increase over time, which brings great challenges to the link prediction task. To address this, we propose a dynamic heterogeneous network link prediction method based on incremental embedding named INE-DHNLP. This method divides the network into multiple subgraphs in term of the type of edges, and performs temporal random walks in different subgraphs to generate random walk sequences with temporal information. Then, the node embedding vectors are incrementally updated for nodes in each subgraph, and use attention mechanisms to aggregate node embedding vectors in different subgraphs. In order to deal with the challenge of node increment affecting link prediction, we designed the dynamically expandable decoder (DED) using incremental learning technology to predict the future link situation of the network. Experimental results on two real datasets show that this method has better comprehensive performance compared to other baseline methods.