Hawkes processes were benchmarked too narrowly: with spline baselines and kernels they match or beat THP, SAHP and NHP on all five standard datasets
Synopsis
Neural temporal point processes, which parameterize event intensities with neural networks, were reported to substantially beat Hawkes processes, but those comparisons gave Hawkes only a constant baseline and a fixed exponential kernel. With learnable spline baselines and kernels plus inhibitive variants, the best Hawkes model matches or exceeds THP, SAHP and NHP on five standard datasets, while a gap to the flexible NeuralSplineTPP and IntensityFree remains and is substantial only on Amazon; the authors also release TorchPP, an open-source framework implementing both families under one interface.
Interpretation
The previously reported gap between Hawkes processes and neural temporal point processes largely disappears once Hawkes processes are given learnable spline baselines and spline triggering kernels. Earlier benchmarks restricted Hawkes processes to a constant baseline and a fixed exponential kernel, then compared them against far more flexible neural models, and omitted health data entirely. On the five standard datasets Taxi, StackOverflow, Retweet, Amazon and Taobao, the best Hawkes variant matches or beats THP, SAHP and NHP in joint log-likelihood; Inhib. Spline/Spline Hawkes is the best model overall on Retweet, Spline/Spline Hawkes ties NeuralSplineTPP on StackOverflow and Taobao, the remaining gap to the flexible neural models is small on Taxi, and only Amazon shows a clear advantage.
The flexibility of the intensity parametrization also limits neural temporal point processes, and the benefit falls almost entirely on the time density of events. Earlier neural models mostly used Softplus-wrapped intensity heads that require numerical integration and restrict the shape in time. NeuralSplineTPP, a Transformer with a spline intensity head, outperforms THP, SAHP, NeuralConstTPP and NHP and is on par with the intensity-free IntensityFree; decomposing the log-likelihood into a time term and a type-conditional term, raising the spline knot count improves the time term by 0.9–1.6 nats on Retweet and Amazon while the type term does not improve on any dataset.
On two large-scale medical event-sequence tasks, pairwise additive interactions already explain most of the next-event structure, and the neural advantage concentrates on events recorded in the same minute. Earlier comparisons did not cover large-scale medical data such as electronic health records. On the 1,257-type first-occurrence disease histories from UK Biobank, Spline/Spline Hawkes trails NeuralSplineTPP by only a small number of nats in joint log-likelihood and outperforms the other neural models; on the 117-type ICU events from MIMIC-IV, NeuralSplineTPP leads the best Hawkes variant, with the gap concentrated on events recorded in the same minute as the previous event and small on the remaining events.
Perspective
The conclusions cover Poisson and Hawkes processes and the evaluated Transformer intensity models, NHP and IntensityFree, on autoregressive next-event prediction over five standard datasets plus the UK Biobank and MIMIC-IV medical tasks; long-horizon simulation, planning and decision making, and Cox-type, flow, diffusion and state-space models are not evaluated.
The authors note that the conclusions rest on seven datasets and that other domains may favor neural models more strongly; the UK Biobank task omits sex and lifestyle covariates and inhibitive Hawkes processes were not evaluated there; MIMIC-IV uses only a subset of recorded values and is not a medically relevant forecasting model; Retweet's integer time stamps let a continuous density reach an unbounded time likelihood, so scoring must use the recorded time bins.
