EventFlow: Forecasting Temporal Point Processes with Flow Matching
This addresses the issue of cascading errors in autoregressive models for event sequence forecasting, offering a more efficient solution for domains like industrial and scientific data analysis.
The paper tackles the problem of forecasting long-horizon event sequences in temporal point processes by proposing EventFlow, a non-autoregressive generative model based on flow matching, which reduces forecast error by 20%-53% compared to baselines.
Continuous-time event sequences, in which events occur at irregular intervals, are ubiquitous across a wide range of industrial and scientific domains. The contemporary modeling paradigm is to treat such data as realizations of a temporal point process, and in machine learning it is common to model temporal point processes in an autoregressive fashion using a neural network. While autoregressive models are successful in predicting the time of a single subsequent event, their performance can degrade when forecasting longer horizons due to cascading errors and myopic predictions. We propose EventFlow, a non-autoregressive generative model for temporal point processes. The model builds on the flow matching framework in order to directly learn joint distributions over event times, side-stepping the autoregressive process. EventFlow is simple to implement and achieves a 20%-53% lower forecast error than the nearest baseline on standard TPP benchmarks while simultaneously using fewer model calls at sampling time.