Publikation
Peculiarities of Counterfactual Point Process Generation
Gerrit Großmann; Sumantrak Mukherjee; Sebastian Vollmer
In: Jianwu Wang; Sahara Ali; Yanan Xin (Hrsg.). 32nd ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems. ACM SIGSPATIAL International Workshop on Spatiotemporal Causal Analysis (STCausal-2024), 1st ACM SIGSPATIAL International Workshop on Spatiotemporal Causal Analysis, located at 32nd ACM SIGSPATIAL, October 29 - November 1, Atlanta, GA, USA, ISBN 979-8-4007-1154-1/24/10, Association for Computing Machinery, New York, NY, United States, 11/2024.
Zusammenfassung
(Spatio-)Temporal point processes are the de facto standard for modeling events in (space and) time. This study addresses the generation of counterfactual sequences: Given a model and an observed sequence, we investigate how the sequence would have evolved under a hypothetical alternative model.
Our contribution is two-fold: Firstly, we demonstrate how to easily leverage established stochastic simulation algorithms to generate counterfactual sequences. Secondly, we reveal that different simulation methods---despite being statistically equivalent---correspond to distinct structural causal models of the point process, producing distinct counterfactual distributions.
Given these findings, we recommend exercising greater caution when applying counterfactual reasoning in this domain, particularly concerning the relationship between counterfactual generation and the underlying physical processes.