We are excited to announce that our paper "Learning Conditional Marked Event Sequences with Mixed Data Types" was accepted at NeurIPS 2025 as spotlight (top ~3% of submissions). This paper was a result of the HPI-UCI collaboration.
Authors
Felix Draxler, Yang Meng, Kai Nelson, Lukas Laskowski, Yibo Yang, Theofanis Karaletsos, Stephan Mandt
Abstract
Marked Temporal Point Processes (MTPPs) model sequences of irregularly timed events accompanied with metadata (marks), which occur in various domains such as medicine, finance, and remote sensing. We extend MTPPs in two ways: First, we support variable-length, mixed-type (i.e., both discrete and continuous) marks, such as those found in electronic health records. Second, we treat conditioning as a first-class modeling principle by using our flexible MTPPs as the structured output of a regression task such as the detection of events in an input time series. Our model uses an autoregressive Transformer to directly model the joint distribution of event times and marks, employing flexible normalizing flow models for continuous-valued marks. Our intensity-free formulation avoids numerical integration and naturally supports complex, conditional event sequences. Empirically, we find that our model excels both at tasks with discrete-only and mixed-type marks, and that the gained flexibility improves prediction quality.
