EventFormer: a hierarchical neural point process framework for spatio-temporal clustering events prediction

Abstract In real-world scenarios, event data often exhibits inherent randomness, complex historical dependencies, and hierarchical spatio-temporal clustering. However, existing neural point process models typically overlook the hierarchical nature of spatial information, or treat temporal and spatial relevance as separate factors. This oversight results in suboptimal performance when handling data with prominent spatio-temporal clustering features. Additionally, current models have not explicitly considered the interdependencies between event types. To remedy these limitations, we introduce a novel neural point process framework named EventFormer. Leveraging the time-oriented and type-oriented multi-head attention modules, along with a Ladder Attention mechanism that progressively refines spatial embeddings across hierarchical levels, EventFormer adeptly captures the nuanced dynamics of event occurrences. Furthermore, EventFormer incorporates a type-aware conditional intensity function to explicitly model interactions between event types, enhancing both predictive accuracy and interpretability. Extensive experiments on real-world datasets demonstrate the outstanding performance of EventFormer in event likelihoods modeling and prediction tasks.

Abstract licence: CC BY-NC-ND

1. 7.

   Attention Is All You Need · 2017

41 more references, whose wording the paper’s licence keeps private.

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.