FeatHawkes: Scalable Feature-Based Attribution for Temporal Event Data

ICLR 2026 Conference Submission12940 Authors

Keywords: Attribution, Hawkes Process, Methodology, Sport Analytics

Abstract: Attribution, the problem of assigning proportional responsibility for an outcome to each event in a temporal sequence of causes, is central to diverse applications ranging from marketing and seismology to sports analytics. While incorporating exogenous features substantially enhances the expressiveness of attribution models, existing approaches lack the scalability required to integrate modern machine learning methodology. We introduce FeatHawkes, a feature-augmented Hawkes process framework for event-level attribution in continuous time. Our core contribution is a novel first-order optimization routine for Hawkes processes that leverages stochastic gradient methods, scaling favorably with both dataset size and feature dimensionality. This gradient-based formulation enables compatibility with automatic differentiation and end-to-end ML pipelines. We release FeatHawkes as an open-source Python library, and demonstrate its effectiveness through synthetic experiments and a case study on professional football data, where the framework supports what-if analyses such as quantifying the impact of replacing players in a lineup.

Primary Area: learning on time series and dynamical systems

Submission Number: 12940