STGEN: Deep Continuous-Time Spatiotemporal Graph Generation

Spatiotemporal graph generation has realistic social significance since it unscrambles the underlying distribution of spatio-temporal graphs from another perspective and fuels substantial spatio-temporal data mining tasks. Generative models for temporal and spatial networks respectively cannot be easily generalized to spatiotemporal graph generation due to their incapability of capturing: 1) mutually influenced graph and spatiotemporal distribution, 2) spatiotemporal-validity constraints, and 3) characteristics of multi-modal spatiotemporal properties. To this end, we propose a generic and end-to-end jointly captures the graph, temporal, and spatial distributions of spatiotemporal graphs. Particularly, STGEN learns the multi-modal distribution of spatiotemporal graphs via learning the distribution of spatiotemporal walks based on a new heterogeneous probabilistic sequential model. Auxiliary activation layers are proposed to retain the spatiotemporal validity of the generated graphs. In addition, a new boosted strategy for the ensemble of discriminators is proposed to distinguish the generated and real spatiotemporal walks from multi-dimensions and capture the combinatorial patterns among them. Finally, extensive experiments are conducted on both synthetic/real-world spatio-temporal graphs and demonstrated the efficacy of the proposed model.