Add to Chrome
Add to Firefox
Add to EdgeIntensity-free Integral-based Learning of Marked Temporal Point Processes
Aug 07, 2023
In the marked temporal point processes (MTPP), a core problem is to parameterize the conditional joint PDF (probability distribution function) $p^*(m,t)$ for inter-event time $t$ and mark $m$, conditioned on the history. The majority of existing studies predefine intensity functions. Their utility is challenged by specifying the intensity function's proper form, which is critical to balance expressiveness and processing efficiency. Recently, there are studies moving away from predefining the intensity function -- one models $p^*(t)$ and $p^*(m)$ separately, while the other focuses on temporal point processes (TPPs), which do not consider marks. This study aims to develop high-fidelity $p^*(m,t)$ for discrete events where the event marks are either categorical or numeric in a multi-dimensional continuous space. We propose a solution framework IFIB (\underline{I}ntensity-\underline{f}ree \underline{I}ntegral-\underline{b}ased process) that models conditional joint PDF $p^*(m,t)$ directly without intensity functions. It remarkably simplifies the process to compel the essential mathematical restrictions. We show the desired properties of IFIB and the superior experimental results of IFIB on real-world and synthetic datasets. The code is available at \url{https://github.com/StepinSilence/IFIB}.
