Event Time Prediction via Survival Analysis Based Multivariate Transformer Point Process

Yan Zhou, Xiaodong Li

Abstract: Events profoundly impact various facets of our society, highlighting the critical significance of event time prediction. In practical scenarios, events mostly continue for some time, and their duration is crucial for in-depth mining. Although Neural Temporal Point Process (NTPP) models have gained widespread acclaim for modeling event sequences, existing NTPP models typically perceive events as discrete points, which makes them predict only the occurrence time of events. To tackle this limitation, we introduce the Survival Analysis based Multivariate Transformer Point Process (SAMTPP) model. Specifically, we design a multivariate intensity function to concurrently capture events’ occurrence and end time. Given that this multivariate function introduces computational complexities, we facilitate some necessary decompositions through survival analysis techniques. Moreover, SAMTPP is built based on the Transformer architecture to capture intricate event interactions, enabling end-to-end predictions of the occurrence and end time. Empirical experiments conducted on several real-world financial datasets demonstrate the effectiveness of SAMTPP in efficiently forecasting both event times, outperforming all benchmark models. To the best of our knowledge, SAMTPP is the first NTPP model capable of simultaneously predicting both the occurrence and end time of events.

External IDs:dblp:conf/dasfaa/ZhouL24