Continuous temporal network embedding by modeling neighborhood propagation process

Network embedding has become the focal point of increasing research interests in academic and industrial domains. Temporal networks are evolved by adding nodes and edges sequentially, which should be regarded as neighborhood propagation processes driven by chronological interactive events between nodes at arbitrary time granularity. In this scenario, both the occurrence times and connected nodes of interactive events are the key factors for embedding. Existing models of neighborhood propagation process focus on conventional temporal point process modeled by manual parametric forms of conditional intensity function, which merely measures the event occurrence probability at occurrence time. However, modeling event propagation without considering the connections of nodes in the network is inaccurate, requiring further research to capture both occurrence times and target nodes for temporal network embedding. In this paper, we propose a point process-based continuous temporal network embedding method by modeling neighborhood propagation process, named NPPCTNE, where the neighborhood propagation history is embedded into a hidden vector to jointly model the node and time of each interactive event. For propagation node prediction, the generating process of the next target node is simulated respectively by transient and steady-state responses to aggregate the long-term and short-term propagation effects of the continuous network system. For propagation time prediction, an intensity-free temporal point process is applied, which directly estimates the generative time probability distribution of the propagation sequence to learn the whole evolution process of the temporal network. Experiments on four real-world datasets show the efficacy of our model compared to state-of-the-arts.