Link prediction method for heterogeneous dynamic network

The invention discloses a link prediction method for a heterogeneous dynamic network, and the method comprises the following steps: S1, generating a time sequence sub-graph, determining a node type and an edge type which are used for constructing the time sequence sub-graph in a heterogeneous data set, selecting a timestamp, and determining the structure of the time sequence sub-graph at each moment; s2, network embedding learning is carried out, historical interaction information of nodes and structure information of the time sequence sub-graphs are fused, and vector representation of each time sequence sub-graph is obtained; s3, link prediction: processing graph data by adopting an LSTM network, taking the vector representation of the time sequence sub-graph as the input of the LSTM, capturing time sequence information, and calculating the similarity of node pairs according to an output hidden state vector so as to measure the probability of interaction of the node pairs; and S4, model optimization: a link