Automatic Integration for Neural Temporal Point Process

Zihao Zhou, Rose Yu

Keywords: point process

Abstract: Integration lies at the heart of the temporal point process. Due to the intrinsic mathematical difficulty of symbolic integration, neural temporal point process models either constrain the intensity function to an integrable functional form or apply certain numerical methods. However, the former type of model has limited expressive power, and the latter type of model suffers additional numerical errors and high computational costs. In this paper, we introduce automatic integration with neural point process models, a new paradigm for efficient, closed-form nonparametric inference of temporal point process characterized by any intensity function. We test the model against a variety of synthetic temporal point process datasets and show that the model can better capture inter-event intensity changes than state-of-the-art methods. We also identify certain model settings that would lead the MLE estimator for the temporal point process to be inconsistent.

One-sentence Summary: Leveraging automatic integration for more efficient and accurate recovery of temporal point process's intensity