Abstract
We present a novel Neural Embedding Spatio-Temporal (NEST) point processmodel for spatio-temporal discrete event data and develop an efficientimitation learning (a type of reinforcement learning) based approach for modelfitting. Despite the rapid development of one-dimensional temporal pointprocesses for discrete event data, the study of spatial-temporal aspects ofsuch data is relatively scarce. Our model captures complex spatio-temporaldependence between discrete events by carefully design a mixture ofheterogeneous Gaussian diffusion kernels, whose parameters are parameterized byneural networks. This is the key that our model can capture intricate spatialdependence patterns and yet still lead to interpretable results as we examinemaps of Gaussian diffusion kernel parameters. The imitation learning modelfitting for NEST is more robust since it directly measures the divergencebetween the empirical distributions between the training data and themodel-generated data. Moreover, our imitation learning-based approach enjoyscomputational efficiency due to the explicit characterization of the rewardfunction related to the likelihood function; furthermore, the likelihoodfunction under our model enjoys tractable expression due to Gaussian kernelparameterization. Experiments based on real data show our method's goodperformance relative to the state-of-the-art and the good interpretability ofNEST's result.