Automatic Integration for Fast and Interpretable Neural Point Processes

Abstract

The fundamental bottleneck of learning continuous-time point processes is integration. Due to the intrinsic mathematical difficulty of symbolic integration, neural point process (NPP) models either constrain the intensity function to a simple integrable kernel function or apply numerical integration. However, the former has limited expressive power. The latter suffers additional numerical errors and high computational costs. In this paper, we introduce *Automatic Integration for Neural Point Process* models (Auto-NPP), a new paradigm for exact, efficient, non-parametric inference of point process. We validate our method on simulated events governed by temporal point processes and real-world events. We demonstrate that our method has clear advantages in recovering complex intensity functions from irregular time series. On real-world datasets with noise and unknown intensity functions, our method is also much faster than state-of-the-art NPP models with comparable prediction accuracy.