B1476
Title: Learning quantile functions for temporal point processes with recurrent neural splines
Authors: Souhaib Ben Taieb - University of Mons (Belgium) [presenting]
Abstract: The combination of temporal point process (TPP) models with deep neural networks provides a powerful and flexible framework for modeling continuous-time event data. Neural TPP models are autoregressive models which characterize the distribution of the next arrival time conditional on the observed history. Various representations of this conditional distribution have been proposed in the literature, including parametric forms for the intensity function, the density function, or the cumulative intensity function. Which function to parametrize and how to parametrize it is an important design choice. We propose a new recurrent neural TPP model which parametrizes the conditional quantile function of the inter-arrival times with a monotonic rational-quadratic spline. While being flexible, our spline-based parameterization has closed-form expressions for multiple useful quantities such as the expectation, the likelihood function, or any quantile. We also derive a closed-form expression for the Continuous Ranked Probability Score (CRPS) which enables efficient model optimization. Finally, we demonstrate that the proposed model achieves competitive performance compared to state-of-the-art neural TPP models on both synthetic and real-world event sequence data.
