A NOVEL APPROACH FOR LEARNING TEMPORAL POINT PROCESS

Abstract

In this paper, we presented a novel methodology for learning temporal point process based on the implementation of one-dimensional numerical integration techniques. The implementation of numerical methodology is used for linearizing negative maximum likelihood (neML) function to enable backpropagation of neML derivative. The presented approach is tested on highway toll dataset. Moreover, four different wellknown point process baseline models were compared: first-order and second-order polynomial Poisson inhomogeneous process and Hawkes with exponential and Gaussian kernel. The results showed that different numerical integration techniques influence the quality of the obtained models.