Non-autoregressive diffusion-based temporal point processes for continuous-time long-term event prediction

Wang-Tao Zhou

doi:10.1016/j.eswa.2024.126210

Non-autoregressive diffusion-based temporal point processes for continuous-time long-term event prediction

2025-04-01

0

PRE

AI

DOI：10.1016/j.eswa.2024.126210

Original

Original request for help

Save

Abstract

En 中文

Continuous-time event prediction plays an important role in many applications, especially for time-sensitive scenarios. Temporal Point Processes (TPPs) are handy tools for this task. However, conventional TPP models rely on autoregressive prediction and ignore the internal consistency of the predicted event sequence, which suffers from error accumulation in the long term, compromising the ability to anticipate events in the distant future accurately. In this work, we propose a diffusion-based TPP framework for long-term event prediction in continuous time. The framework has an auto-encoder-like structure, supporting efficient end-to-end training and robust non-autoregressive sampling. We further devise a novel denoising architecture tailored for the setting of long-term event prediction, which jointly considers history dependency and internal consistency of the target sequence to provide a high-quality denoising signal. Extensive experiments are conducted to prove the superiority of our proposed model over state-of-the-art methods. To the best of our knowledge, this is the first work to apply diffusion methods to long-term event prediction problems.

Keywords:

Generative model

Denoising diffusion probabilistic model (DDPM)

Deep learning

Residual network

Dilated convolution

Multi-step prediction

Non-autoregressive diffusion-based temporal point processes for continuous-time long-term event prediction

Abstract

Journal

Organization