B1557
Title: Temporal event boosting: Gradient boosting applied to conditional intensity models
Authors: Fredrik Lundvall Wollbraaten - University of Oslo (Norway) [presenting]
Abstract: Temporal point process data arise in many real-world settings due to increased focus on capturing and storing data across many fields. We consider Marked Temporal Point Processes (MTPP), where each arrival time $t_i \in \mathbb{R}^+$ has a discrete mark $m_i \in \{1, ..., M\}$, which we refer to as an event type. Typical examples are system log messages, neural spiking activity, online customer behavior data, events in football matches, or any other setting where discrete events occur in continuous time. Considering the MTPP as a multivariate point process, we propose Temporal Event Boosting (TEB) for estimating mark-specific conditional intensities depending on the history. Despite the success of gradient boosting, its extension to MTPP has not been considered. TEB is a gradient boosting approach for MTPP based on discretizing time, encoding the history of the process using counts of events in different intervals, whereafter gradient boosting is applied. The method is simple to implement using existing software. Using both simulated and real data (trading and football data), we show that TEB performs very well compared to parametric and recurrent neural network-based alternatives.
