Latent Block-Diffusion Temporal Point Processes: A Semi-Autoregressive Framework for Asynchronous Event Sequence Generation

Traditional Temporal Point Processes (TPPs) face a dilemma: autoregressive models generate sequences one event at a time, leading to significant error accumulation over long horizons, while non-autoregressive diffusion models generate sequences in parallel but are typically restricted to fixed-length outputs. This paper asks whether a hybrid, semi-autoregressive approach can combine the flexibility of the former with the parallel generation quality of the latter.

The authors propose Latent Block-Diffusion Temporal Point Processes (LBDTPP). The framework operates in three stages:

1. Latent Mapping: Raw events (timestamps and marks) are embedded into a continuous latent space to create a homogeneous representation.
2. Block-wise Autoregression: The latent sequence is partitioned into blocks. The model generates these blocks sequentially, preserving the long-term dependencies and variable-length nature of the sequence.
3. Intra-block Diffusion: Within each block, the model uses Gaussian diffusion to generate multiple event representations in parallel. This reduces the number of sequential steps, thereby limiting the propagation of errors.

The model is trained by minimizing a combination of a latent block diffusion loss and a reconstruction loss, which maps the latent representations back to the original event space.

Theoretical analysis using Wasserstein error bounds demonstrates that LBDTPP reduces error accumulation by shortening the recursive sampling horizon from individual events to blocks. Empirical experiments across six real-world datasets show that LBDTPP consistently outperforms state-of-the-art autoregressive and non-autoregressive baselines in both unconditional and conditional generation tasks. Furthermore, the model maintains competitive inference speeds, with a fast-sampling variant that can outperform existing methods in efficiency.