LBDTPP: Semi-Autoregressive Framework for Asynchronous Event Sequence Generation

Modeling and generating asynchronous event sequences, common in areas like social networks, medical diagnoses, and financial transactions, is a critical task. Existing autoregressive methods often suffer from accumulating errors during multi-step generation, while non-autoregressive diffusion models are typically restricted to producing fixed-length outputs. To address these limitations, researchers have developed Latent Block-Diffusion Temporal Point Processes (LBDTPP). LBDTPP is a novel semi-autoregressive framework that introduces a latent block diffusion mechanism. The core idea involves defining an autoregressive probability distribution over "event blocks" within a latent space. Simultaneously, Gaussian diffusion is performed within each of these blocks. This dual approach allows LBDTPP to sequentially generate blocks while concurrently sampling events within each block, combining the length flexibility of autoregressive models with the high-quality, parallel generation capabilities of diffusion models. The theoretical underpinnings include Wasserstein error bounds, demonstrating that block-wise generation can reduce error accumulation compared to event-wise autoregressive methods under certain conditions. Extensive experiments across six real-world benchmark datasets show that LBDTPP outperforms state-of-the-art Temporal Point Process (TPP) baselines in both unconditional and conditional generation tasks. Further analysis confirms the benefits of latent-space diffusion and block-wise generation, highlighting a trade-off between generation quality and block size.