Continuous-time Dynamic Network Node Classification Algorithm Based on Autoencoder

Node classification is an important issue in network researches. Existing methods lay more emphasis on the classification of static network nodes, while there are insufficient researches on network dynamic characteristics. At the same time, the existing adjacency matrix representation cannot reflect the potential relationship of indirectly connected nodes. To better describe the network structure and utilize network time information, the author proposes a new Node Classification Algorithm based on Autoencoder, namely NCAA. It adopts the random walk in limited spatio-temporal and the matrix representation of node relationship. The low-dimensional representation vector of the node is generated by the autoencoder model, and the LogicRegression classifier is used to classify the representation vectors, thereby achieving node classification. Experiments show that, compared to some classic algorithms, NCAA can better represent the network, and its two-dimensional visualization of node vectors has a superior effect. Under the same conditions, the F1 value of the classification result is significantly higher than that of baseline algorithms.