TNSEIR: A SEIR pattern-based embedding approach for temporal network

Network embedding defines a set of techniques for learning the vector representations of nodes and links, which accurately encode the network in a low-dimensional space for computational network analysis. The states of a network evolution are multiple containing link emergence and vanishing. However, determining the topological change caused by link vanishing in the dynamic environment is difficult because most temporal network embedding methods are restricted by learning frameworks. For example, embedding methods based on graph neural networks and matrix factorization handle only link emergences. To identify the structural changes caused by both link appearance and link vanishing, this paper introduces a temporal network embedding approach named TNSEIR inspired by classical susceptible-exposed-infectious-recovered (SEIR) model of infectious diseases, which exploits the information of network structure and temporal evolution. The structural change after link disappearance is represented as the ”recovered” state of SEIR. A new node pair connection (”infection”) probability function is proposed for capturing the information of neighboring nodes and the effect of macro-temporal factor. The macro-temporal factor specifies the topological structure of each network snapshot and the ”latency” information of node pairs, while ”latency” is derived from the ”exposed” state of SEIR. Through that, TNSEIR can accurately capture the long-distance connection trends of node pairs and thereby predict the evolution trend. Extensive experiments are performed on three real-world temporal networks. TNSEIR mechanism outperforms the state-of-the-art (SOTA) autonomous-modeling embedding methods on network analysis and network inference tasks. The strength of TNSEIR is distinctly evidenced by its large improvement on the Wikipedia network, which contains both link appearances and disappearances.