betaFeedback

WorksComputer Science

Flexible Intensities Matter: A Comprehensive Re-evaluation of Hawkes Processes and Neural Temporal Point Processes

Abstract

Temporal point processes (TPPs) are the canonical model class for event sequences in continuous time, with applications ranging from seismology to electronic health records. Recent benchmarks report that neural TPPs substantially outperform their classical counterparts, such as Hawkes processes. However, these comparisons set flexible neural models against Hawkes processes in their most restricted forms and covered only a narrow range of domains. We revisit this comparison under autoregressive next-event prediction, evaluating Poisson and Hawkes processes of increasing flexibility against neural TPPs on five established benchmarks and two large-scale medical tasks derived from UK Biobank disease histories and MIMIC-IV ICU records. First, once Hawkes processes are equipped with flexible, learnable baselines and kernels, the previously reported gap to neural TPPs largely closes. Second, neural TPPs are likewise limited by their intensity parametrization: a Transformer with a spline-based intensity head outperforms previous intensity-based neural TPPs and matches or exceeds the intensity-free model. Decomposing the log-likelihood into time and type terms shows that this gain lies almost entirely in modeling when events occur (log p(t)) rather than which type occurs (log p(d|t)). Third, only on MIMIC-IV do neural TPPs retain an advantage in predicting event types, underscoring the value of complex datasets for evaluating TPPs. We release TorchPP, an open-source PyTorch framework with unified implementations of classical and neural TPPs.

Introduction

Temporal point processes (TPPs) provide a principled framework for modeling irregular event sequences in continuous time. Applications range from seismology [25] and social media analysis [44] to healthcare, where a patient’s medical history forms a sequence of clinical events [32].

The Hawkes process [8] is a classical TPP that decomposes the intensity into a baseline (exogenous) component and triggering kernels (endogenous) through which past events excite future ones.

In recent years, neural TPPs have gained traction. They replace this additive structure with learned representations: a neural encoder summarizes the event history, and a flexible output head parametrizes either the conditional intensity or the inter-event time distribution [31, 40]. See Figure 1 for a conceptual visualization.

Conceptual comparison of Hawkes processes and neural TPPs

Figure 1. Conceptual comparison of Hawkes processes (1) and neural TPPs (2), illustrated on an event sequence with two event types (• type A, ▲ type B). Top (Hawkes process). Each past event interacts via a kernel response ϕdi→d(t−ti)\phi_{d_i\to d}(t-t_i) (colored curves) on top of a (possibly time-varying) baseline μd(t)\mu_d(t) (thin solid line). The total intensity (upper sub-panel) is the sum of baseline and active kernels. At any query time tt, the intensity is thus the baseline plus the sum of kernel contributions from all past events. Bottom (Neural TPP). A neural encoder processes the same event history and outputs a hidden state at each event location. The head of the neural TPP projects this encoding to an intensity function. Conceptually, this lets neural TPPs capture higher-order, non-additive interactions between events, at the cost of the interpretable decomposition of the Hawkes process.

Preprint.

Neural TPPs have repeatedly been reported to outperform Hawkes baselines [6, 21, 38, 42, 33] (Shchur et al. 2020b), most recently and by substantial margins in the EasyTPP benchmark [40].

However, the Hawkes baselines in these comparisons were typically restricted to a constant baseline and exponential kernels, far less flexible than the neural TPPs they were compared against. These evaluations also cover a narrow range of domains and omit health data entirely, despite growing interest in sequence models for electronic health records [32, 35, 15, 27, 28].

In this work, we systematically re-evaluate Poisson processes, Hawkes processes, and neural TPPs on established benchmark datasets and on two large-scale medical tasks: disease histories from the UK Biobank [Sudlow et al., 2015] and ICU records from MIMIC-IV [11].

To our knowledge, this is the first systematic comparison of Hawkes processes with flexible baselines and kernels and inhibitive variants against modern neural TPPs under a unified protocol. Prior benchmarks [40, 2] considered Hawkes processes with a constant baseline and an exponential kernel [8], while Zhao and Engelhard Zhao and Engelhard [ [43] used a neural but purely excitatory kernel. We deliberately restrict the set of classical models to Poisson and Hawkes processes; Cox-type models such as log-Gaussian Cox processes are beyond the scope of this work. On the neural side, we consider RNN- and Transformer-based intensity models with output heads of increasing flexibility, as well as an intensity-free model. Flow-, diffusion-, and state-space-based TPPs are out of scope. The code is available at https://github.com/PMBio/torchPP, together with checkpoints for the standard benchmarks.

Our contributions are as follows:

  • We show that the previously reported gap between Hawkes processes and neural TPPs is largely an artifact of under-specified baselines: with spline baselines, spline kernels, and an inhibitive variant, Hawkes processes match or exceed the performance of rigidly parametrized neural TPPs (THP, SAHP, NHP) and are competitive with the flexible neural models (IntensityFree, NeuralSplineTPP), indicating that these benchmarks are largely explained by pairwise interactions.

  • We identify the flexibility of the intensity parametrization as a main driver of performance: a Transformer with a spline-based intensity head (NeuralSplineTPP) outperforms softplus-headed intensity models (THP, SAHP, NHP) and performs on par with IntensityFree [Shchur et al., 2020a]. Unlike the softplus-headed models, it has a closed-form cumulative intensity that allows exact sampling, and unlike IntensityFree, its type distribution depends on the event time.

  • We show that more flexible intensities improve the modeling of when events occur, not which type occurs. Decomposing the joint log-likelihood into a time term log⁡p(t)\log p(t) and a time-conditional type term log⁡p(d∣t)\log p(d \mid t), and sweeping the number of spline knots, we find that added flexibility improves log⁡p(t)\log p(t) but leaves log⁡p(d∣t)\log p(d \mid t) essentially unchanged. Across models, too, the standard benchmarks separate them far more on time prediction than on type prediction.

  • We evaluate both model families on two large-scale medical evaluation tasks: a 1,257-type terminating disease-history task derived from the UK Biobank, and a 117-type ICU-event task derived from MIMIC-IV. On UKB, flexible Hawkes processes nearly match the best neural TPP. On MIMIC-IV, the flexible neural TPP opens a gap in both joint log-likelihood and time-conditional accuracy.

  • We release TORCHPP, a modular PyTorch framework for classical and neural TPPs. It supports terminating and right-censored sequences, exact sampling, and both analytical and numerical computation of the cumulative intensity (Appendix § 8).

Background and Related Work

Temporal Point Processes

A multivariate temporal point process over DD event types on the interval [0,T][0,T] is characterized by its conditional intensity function λ∗(t)=(λ1∗(t),…,λD∗(t))\lambda^*(t) = (\lambda^*_1(t), \ldots, \lambda^*_D(t)), where λd∗(t)\lambda^*_d(t) gives the instantaneous rate of type-dd events conditioned on the history Ht={(ti,di):ti<t}\mathcal{H}_t = \{(t_i,d_i) : t_i < t\}. The log-likelihood of an observed sequence decomposes as:

log⁡p(HT)=∑i=1Nlog⁡p(ti,di∣Hti)+log⁡S∗(T−tN)=∑i=1Nlog⁡λdi∗(ti)−∫0T∑d=1Dλd∗(t) dt(1)\log p(\mathcal{H}_T) = \sum_{i=1}^{N} \log p(t_i,d_i \mid\mathcal{H}_{t_i}) + \log S^*(T-t_N) = \sum_{i=1}^{N} \log\lambda^*_{d_i}(t_i) - \int_0^T \sum_{d=1}^{D} \lambda^*_d(t)\,\mathrm{d}t \tag*{(1)}

where the first term rewards high intensity at the observed event times, and the second, the cumulative intensity (CI), penalizes intensity throughout the observation window, accounting for the intervals in which no events occur. See Appendix § 9 for a detailed derivation.

Poisson Processes Poisson processes assume a history-independent intensity λd∗(t)=μd(t)\lambda^*_d(t) = \mu_d(t), where μd\mu_d is a non-negative function of time. The homogeneous Poisson process corresponds to the assumption of a constant intensity function μd(t)=μd\mu_d(t) = \mu_d; more expressive parametrizations, e.g. splines or neural networks, are equally admissible.

Hawkes Processes A Hawkes process [8] corresponds to an intensity function that is parametrized as a sum of a baseline rate and pairwise interactions from past events:

λd∗(t)=μd(t)+∑ti<tϕdi→d(t−ti).(2)\lambda^*_d(t) = \mu_d(t) + \sum_{t_i<t} \phi_{d_i\to d}(t-t_i). \tag*{(2)}

Here, μd(t)\mu_d(t) is the baseline intensity of type dd, i.e. the intensity of a Poisson process, and ϕd′→d(⋅)\phi_{d'\to d}(\cdot) is the triggering kernel from type d′d' to type dd. The classical choices are a constant baseline μd(t)=μd\mu_d(t) = \mu_d, corresponding to a homogeneous Poisson process, and an exponential kernel ϕd′→d(Δt)=αd′dβd′de−βd′dΔt\phi_{d'\to d}(\Delta t) = \alpha_{d'd}\beta_{d'd}e^{-\beta_{d'd}\Delta t}; together, they yield a closed-form cumulative intensity [8].

Neural Temporal Point Processes In their most common form, neural TPPs [18] (Shchur et al. 2021) replace the additive pairwise Hawkes structure with a flexible intensity function parametrized by learned representations. A history encoder maps Ht\mathcal{H}_t to a context vector h(t)h(t), which is projected to an intensity function by an output head. Common choices for encoders include recurrent networks (RMTPP [6], NHP [21]) and attention-based architectures (THP [46], SAHP [42], AttNHP [41]). Most output heads apply a clamping function such as Softplus to enforce non-negativity, which precludes the closed-form calculation of the cumulative intensity and hence requires numerical integration during training and evaluation. Alternatively, the cumulative intensity itself can be parametrized by a monotone network [45] (Omi et al. 2019). Other approaches specify the distribution of the next event directly, e.g. via mixture models [29] or triangular maps [30], or generate whole sequences via flow matching [13, 20] or diffusion [17, 19, 22]. Neural ODEs [34] and state-space models [5] are further families. Of these, we include the intensity-free mixture model IntensityFree [29] as a strong density-based baseline. The flow, diffusion, and state-space families remain complementary to the intensity/density-flexibility axis studied here and are left to future work.

See Shchur et al. [2021] for a survey of neural TPP architectures, and Figure 1 for a conceptual comparison with Hawkes processes. While neural TPPs have the capacity to learn higher-order interactions beyond pairwise kernels, they sacrifice the intrinsic interpretability of classical methods such as Hawkes processes.

Existing Benchmarks of TPPs

Few benchmarks have systematically compared different neural TPPs. Among these, EasyTPP [40] is widely used and has evaluated several recent TPP architectures across five standard datasets (see § 3.2). However, their Hawkes baseline is highly restricted: it uses a constant baseline intensity, exponential kernels with fixed, non-learned decay rates, and no inhibition. These choices severely limit its expressiveness (see Appendix § 8.§ 6 for a reproduction). Bosser and Ben Taieb [2023] benchmark combinations of neural encoders and decoders, including an exponential-kernel Hawkes decoder, and Lin et al. [2025] compare neural TPP components empirically. Neither considers baselines with flexible Hawkes kernels.

Another related benchmark is Jiang et al. [2025], a multi-modal TPP benchmark for text and video data, which, however, targets a different domain and model class. Earlier work (Du et al. [2016], Mei and Eisner, 2017, Xiao et al., 2017, Zhang et al., 2020, Shchur et al., 2020b, Soen et al., 2021) has also compared neural TPPs with Hawkes processes. As above, these comparisons used only exponential-kernel Hawkes processes [8] or self-correcting processes [9], both of which have a fixed parametric form.

Evaluation Design

Considered Methods

We systematically compare Hawkes processes and neural TPPs with increasingly flexible intensity parametrizations, under a unified training, tuning, and evaluation protocol, implemented in our TorchTPP framework. For full model details see Appendix § 10.

Hawkes Processes

Previous TPP benchmarks have largely considered Hawkes baselines restricted to a constant baseline and an exponential kernel. We consider a more expressive family of Hawkes processes with flexible baseline intensities and kernels, including variants that capture inhibitive effects.

Baseline intensity. For the per-type baseline intensity μd(t)\mu_d(t), we consider two parametrizations:

  1. a constant baseline μd(t)=μd\mu_d(t) = \mu_d (homogeneous Poisson),

  1. a linear spline over time since sequence start, with learnable knot heights (spline Poisson), which captures non-stationary intensities; the number and spacing of the knots are hyperparameters.

Triggering kernel. For the kernels ϕd′→d(Δt)\phi_{d'\to d}(\Delta t), we compare:

  1. the exponential kernel ϕd′→d(Δt)=αd′dβd′de−βd′dΔt\phi_{d'\to d}(\Delta t)=\alpha_{d'd}\beta_{d'd}e^{-\beta_{d'd}\Delta t} (exponential kernel),

  2. a learnable linear spline over the time lag Δt\Delta t for each type pair (d′,d)(d',d) (spline kernel), constrained to be non-negative via Softplus in the non-inhibitive models.

Both kernels admit closed-form cumulative intensities.

Based on these triggering kernels, we consider the following Hawkes processes with progressively more flexible baseline intensities and kernels:

  1. constant baseline + exponential kernel →\to Const./Exp Hawkes

  2. constant baseline + spline kernel →\to Const./Spline Hawkes

  3. spline baseline + exponential kernel →\to Spline/Exp Hawkes

  4. spline baseline + spline kernel →\to Spline/Spline Hawkes

Inhibitive variants. The standard Hawkes intensity (Equation (2)) cannot model inhibitive interactions because intensities, as instantaneous rates, must remain non-negative. To capture inhibitive effects, we wrap the sum of baseline and kernel contributions in a non-negative clamping function [3]. We compared ReLU and Softplus, and found Softplus more stable and better-performing. The resulting models lack a closed-form cumulative intensity. We estimate it numerically with a stochastic trapezoidal rule. We report the Inhib. Const./Exp Hawkes and Inhib. Spline/Spline Hawkes variants.

Neural TPPs

As our primary aim is a meaningful comparison of Hawkes processes and neural TPPs, we select a representative set of neural TPPs based on the EasyTPP benchmark [40], which allows direct comparison with previously reported results. For a comprehensive comparison of alternative neural TPP architectures, we refer the reader to the EasyTPP paper. The neural TPPs we consider share a common structure: a neural history encoder enc\mathrm{enc} that maps Ht\mathcal{H}_t to a context vector hth_t, and an output head f(t,ht)f(t,h_t) that maps this context to per-type intensity functions:

λd∗(t)=f(t,enc(Ht)).\lambda_d^*(t)=f(t,\mathrm{enc}(\mathcal{H}_t)).

Most of the models we consider use a Transformer encoder with a joint time/type event embedding and differ mainly in the design of the output head ff, which maps the hidden state to the conditional intensity (or, for IntensityFree, to the inter-event time distribution). Hyperparameters for all models are tuned on the validation set (Appendix D). We also include two models with a different structure: NHP, which uses a continuous-time recurrent encoder, and IntensityFree, which models the inter-event time distribution directly rather than the intensity. Mirroring the Hawkes variants above, we list the models in order of increasing flexibility:

  • NeuralConstTPP: the head predicts a history-dependent but time-constant intensity λd∗(t)=cd(Ht)\lambda_d^*(t)=c_d(\mathcal{H}_t), yielding an analytical CI. This model corresponds to the simplest intensity parametrization that does not allow for time-varying effects.

  • THP / SAHP: re-implementations of the models from [46, 42], serving as representatives of common but inflexible neural TPP architectures. Both models employ a Softplus-wrapped intensity function on top of the encoder state, requiring numerical integration of the cumulative intensity. For a comparison with the original implementations, see Appendix A.5.

  • NHP [21]: a continuous-time LSTM encoder with a softplus intensity head. It keeps the numerically integrated softplus head of THP/SAHP but replaces the attention encoder with a recurrent one, probing the effect of the encoder architecture.

  • NeuralSplineTPP: to match the kernel flexibility introduced for the Hawkes family, we equip the Transformer encoder with a spline-based intensity head. The head predicts per-type linear splines parametrized by knot heights. Unlike SAHP and THP, this yields a closed-form cumulative intensity and exact sampling by inversion, while remaining arbitrarily flexible given enough knots. A similar but piecewise-constant parametrization is used in MOTOR [35].

  • IntensityFree [29]: an intensity-free model that parametrizes the inter-event-time distribution directly as a log-normal mixture, giving a closed-form likelihood with no numerical integration. In our default implementation it uses a Transformer encoder like the intensity models (the GRU encoder of the original model is compared in Appendix § 15).

Overall, we cover attention- and RNN-based intensity and intensity-free models with heads of increasing flexibility (NeuralConstTPP, NHP, THP, SAHP, IntensityFree, NeuralSplineTPP). Our goal is not an exhaustive benchmark of neural TPP architectures, but a targeted test of whether the flexibility of the intensity parametrization explains the previously reported gap to Hawkes processes. We therefore vary this flexibility while keeping the encoder fixed wherever possible. Further neural model families can be readily added to TORCHPP, and we leave their evaluation to future work.

Datasets and Evaluation Tasks

Following the EasyTPP benchmark [40], we evaluate all the considered models on the Taxi [36], Retweet [44], StackOverflow [14], Taobao [39], and Amazon [23] datasets (Table 1).

DatasetDDTerm.Train / Val / TestTotal EventsMean Len
Taxi10No1,400 / 200 / 40074K37.0
StackOverflow22No1,401 / 401 / 401143K64.6
Retweet3No9,000 / 1,535 / 1,520494K41.1
Amazon16No6,454 / 922 / 1,851413K44.7
Taobao17No1,300 / 200 / 500113K56.7
UKB1,257Yes351,656 / 100,473 / 50,2386.2M12.4
MIMIC-IV (ICU, trunc.)117No67,320 / 8,474 / 8,5317.96M94.4

Table 1. Datasets overview. Small- to medium-sized, low-dimensional standard benchmarks are shown in the top rows, and the medical evaluation tasks, with orders of magnitude more instances and event types, in the bottom rows. “D”: number of event types. “Term.”: whether each type occurs at most once. “Mean Len”: mean number of events per sequence. Sequences are ICU-stay-level (with each patient only in one split) for MIMIC-IV and patient-level for UKB.

Disease event sequences from the UK Biobank. Following Shmatko et al. [2025], we consider longitudinal disease histories for more than 500,000 UK Biobank participants [Sudlow et al., 2015]: first-occurrence dates for 1,256 ICD-10-derived diagnosis categories and death, recorded in days since birth (mean sequence duration ≈ 62 years). Details on data preparation and access are provided in Appendix § 13.

Intensive-Care-Unit (ICU) event series data from the MIMIC-IV dataset. We evaluate our models on a task derived from the MIMIC-IV ICU dataset [11], containing event sequences (procedures, medications, and similar) from over 58,000 patients. We use the input and procedure events of each ICU stay, keep the 116 most frequent item types plus an OTHER class (D=117D = 117), and right-truncate sequences to 316 events, the 90th percentile of sequence lengths. Event times are recorded to the minute, and we dequantize the events of each recorded minute uniformly within that minute. For more details see Appendix § 14.

Training and Evaluation Details

Model hyperparameters for each model–dataset combination are selected on the validation-set log-likelihood, employing a multi-stage search (Appendix § 11). The tuning budget per model family is compared in Table 6. Each model is trained with five random seeds (training configuration in Appendix § 11.§ 4), and we report the mean and standard deviation over seeds. All models are trained and selected on the full-sequence log-likelihood (Equation (1)). For comparability with EasyTPP [40], test metrics are reported per event and conditioned on the first event of each sequence, excluding the survival term after the last event (Appendix § 8.§ 4). We report:

FamilyConfigs / datasetSwept hyperparametersIntegrationParams (MIMIC-IV)
Poisson, Const./Exp Hawkes2–closed form117 – 27,495
Spline Poisson, flexible Hawkes≈ 16KbK_b, Kk×TfK_k \times T_f; KK validation-selectedclosed form2,340 – 285,481
Inhibitive Hawkes2inherited from analytical twin, clamping functionnumerical (grid)27,495 – 285,481
NeuralConstTPP, NeuralSplineTPP≈ 20L×H×dL \times H \times d; Kk×TfK_k \times T_f; KK validation-selectedclosed form216,554 – 286,052
THP, SAHP–architecture from EasyTPP, validatednumerical (per interval)216,905 / 231,764
NHP, IntensityFree7encoder (GRU/Transformer), LL, dd, learning ratenumerical / closed form73,066 / 64,746

Table 6. Tuning budget per model family on the standard benchmarks (search configurations per dataset of the systematic grid), integration method, and trainable parameters on MIMIC-IV (D=117D = 117) at the final configuration. Early stopping (patience 4 on the standard benchmarks, 6 on UKB/MIMIC-IV, best validation checkpoint restored) is identical for all families. KbK_b, KkK_k: baseline and kernel knots; TfT_f: kernel support in mean inter-event times; LL, HH, dd: Transformer layers, heads, embedding dimension.

  • the log-likelihood per event, decomposed as log⁡p(t,d∣H)=log⁡p(t∣H)+log⁡p(d∣t,H)\log p(t,d \mid\mathcal{H}) = \log p(t \mid\mathcal{H}) + \log p(d \mid t,\mathcal{H}) into a time term and a type-conditional term;

  • the time-conditional accuracy Acc⁡∣t\operatorname{Acc}_{\mid t}: the top-1 accuracy of p(d∣t,H)p(d \mid t,\mathcal{H}) at the true next-event time tt;

  • the time RMSE of the next-event time prediction E[t∣H]\mathbb{E}[t \mid\mathcal{H}].

Additional metrics are reported in Appendix § 15: the marginal type log-likelihood log⁡p(d∣H)=log⁡∫p(t,d∣H) dt\log p(d \mid\mathcal{H}) = \log\int p(t,d \mid\mathcal{H})\,\mathrm{d}t and the corresponding marginal top-1 accuracy.

The quantities p(d∣H)p(d \mid\mathcal{H}) and E[t∣H]\mathbb{E}[t \mid\mathcal{H}] require numerical integration over the next-event time. Because this is expensive, all test metrics are computed on the first 1,000 test sequences of each dataset (all test sequences for Taxi, StackOverflow, and Taobao) and, for the integrated metrics on MIMIC-IV, on the first 500. See Appendix § 9 for full details.

Results

Revisiting Existing Benchmark Datasets

Table 2 reports the per-event test log-likelihood on the five standard benchmark datasets, split into the time term log⁡p(t)\log p(t) and the type term log⁡p(d∣t)\log p(d \mid t), which sum to the joint log-likelihood. All further metrics are presented in Appendix § 15.

ModelTaxiStackOv.RetweetAmazonTaobao
Poisson processes
H. Poisson0.51/−1.14−0.86/−1.90−3.58/−0.82−0.36/−2.242.07/−2.18
S. Poisson0.51/−1.13−0.85/−1.88−3.43/−0.82−0.35/−2.242.11/−2.19
Hawkes processes
C/E Hawkes0.45/−0.74−0.76/−1.62−3.42/−0.79−0.10/−2.032.61/−1.42
S/E Hawkes0.45/−0.74−0.75/−1.60−3.38/−0.78−0.11/−2.042.61/−1.43
C/S Hawkes0.50/−0.71−0.73/−1.60−3.40/−0.790.01/−1.962.65/−1.38
S/S Hawkes0.49/−0.70__−0.73/−1.59__−2.37/−0.801.73/−2.042.67/−1.41
Inhibitive Hawkes processes
I. C/E Hawkes0.54/−0.33−0.75/−1.60−3.41/−0.780.05/−2.062.62/−1.43
I. S/S Hawkes0.68/−0.28−0.76/−1.60__−1.29/−0.80__2.05/−2.172.66/−1.47
Neural TPPs
N. Const0.55/−0.29−0.80/−1.59−3.42/−0.78−0.34/−1.992.07/−1.43
THP0.56/−0.26−0.80/−1.59−3.41/−0.78−0.31/−1.992.11/−1.45
SAHP0.64/−0.24−0.75/−1.57−3.38/−0.790.14/−2.202.66/−1.49
NHP0.60/−0.24−0.79/−1.60−3.51/−0.81−0.32/−2.062.56/−1.51
Flexible neural TPPs
IntensityFree0.73/−0.29−0.73/−1.60−3.66/−0.792.41/−2.042.73/−1.42
N. Spline__0.71/−0.22__−0.72/−1.60−1.86/−0.782.56/−1.912.71/−1.43

Table 2. Main results: per-event test log-likelihood on the standard benchmark datasets (nats, higher is better; mean over 5 seeds; EasyTPP convention, i.e. conditioned on the first event and without the final survival term, § 3.3). Each cell is log⁡p(t)\log p(t) / log⁡p(d∣t)\log p(d \mid t) (time / type), and the joint log⁡p(t,d)=log⁡p(t)+log⁡p(d∣t)\log p(t,d) = \log p(t) + \log p(d \mid t) is their sum (Table 8; ±std in Appendix § 15). Models differ chiefly in the time term log⁡p(t)\log p(t), the axis that flexibility improves, while the type term log⁡p(d∣t)\log p(d \mid t) is comparatively tight across history-aware models (near-flat on StackOverflow and Retweet). Row labels: baseline/kernel (C: constant, S: spline, E: exponential), I.: inhibitive, N.: neural. Bold: best per dataset and component; underline: best joint log-likelihood; ties at the displayed precision are all marked.

Hawkes processes close the gap. Once Hawkes processes are equipped with flexible baselines, flexible kernels, and inhibitive effects, the previously reported performance gap to neural TPPs narrows substantially (Table 8). On every dataset, the best Hawkes variant matches or outperforms the softplus-headed neural TPPs (THP, SAHP, NHP). Against the flexible neural TPPs, Hawkes models are competitive on four of the five datasets: Inhib. Spline/Spline Hawkes is the best model overall on Retweet, Spline/Spline Hawkes ties with NeuralSplineTPP on StackOverflow and Taobao, and the remaining gap on Taxi is small (0.09 nats). Amazon is the only dataset with a substantial gap (0.77 nats). The performance gap reported in prior benchmarks on these datasets therefore largely reflects under-specified Hawkes models rather than an intrinsic advantage of neural architectures. The relevance of individual model components varies across datasets. A non-stationary baseline intensity (Spline Poisson vs. homogeneous Poisson) provided the biggest benefits on Retweet. Inhibition helped on Taxi, Retweet, and Amazon, but offered no gain over the best Hawkes model on StackOverflow, and reduced performance on Taobao. We attribute these losses to optimization difficulties: numerical integration of the cumulative intensity likely leads to noisier gradients and less stable training. Replacing the exponential kernel with a spline kernel improved the spline-baseline Hawkes process on all datasets, to varying degrees.

ModelTaxiStackOverflowRetweetAmazonTaobao
Poisson processes
Homogeneous Poisson−0.63±.00−2.76±.00−4.41±.00−2.60±.00−0.11±.00
Spline Poison−0.62±.00−2.73±.00−4.25±.00−2.59±.00−0.07±.00
Hawkes processes
Const./Exp Hawkes−0.29±.00−2.38±.00−4.20±.00−2.13±.021.19±.00
Spline/Exp Hawkes−0.29±.00−2.35±.01−4.16±.00−2.15±.011.19±.00
Const./Spline Hawkes−0.21±.00−2.34±.00−4.19±.00−1.95±.001.27±.00
Spline/Spline Hawkes−0.21±.00−2.32±.01−3.17±.00−0.31±.001.27±.00
Inhibitive Hawkes processes
Inhib. Const./Exp Hawkes0.21±.00−2.35±.00−4.19±.00−2.01±.021.18±.00
Inhib. Spline/Spline Hawkes0.40±.00−2.35±.00−2.09±.02−0.12±.011.19±.01
Neural TPPs
NeuralConstTPP0.27±.00−2.40±.01−4.21±.01−2.33±.000.64±.01
THP0.30±.00−2.39±.01−4.19±.02−2.31±.010.66±.01
SAHP0.40±.02−2.32±.01−4.17±.03−2.05±.061.17±.05
NHP0.36±.04−2.40±.02−4.32±.06−2.38±.041.06±.01
Flexible neural TPPs
IntensityFree0.44±.00−2.33±.02−4.45±.260.37±.041.30±.01
NeuralSplineTPP0.49±.00−2.32±.01−2.64±.010.65±.011.28±.01

Table 8. Joint test log-likelihood per event log⁡p(t,d∣H)\log p(t,d \mid\mathcal{H}) (nats, higher is better) on the standard benchmarks. It is the sum of Tables 9 and 10. Mean ±\pm std over 5 seeds. Bold: best per dataset; underline: best classical (Poisson or Hawkes) model; ties at the displayed precision are all marked.

Flexible intensity parametrization matters. Flexible intensity parametrizations benefit neural TPPs as well. NeuralSplineTPP outperforms THP, SAHP, NeuralConstTPP, and NHP by a considerable margin on all datasets except StackOverflow, where it ties with SAHP. As the Transformer models differ mainly in their output head, this is consistent with the flexible but analytically tractable spline head being the decisive factor. The IntensityFree model, which uses a flexible next-event-time density, performs best on Taobao and second best on Taxi and Amazon, but falls behind on Retweet. Notably, the constant-intensity NeuralConstTPP performs comparably to THP, while requiring no numerical integration. Equipped with the GRU encoder of the original model, IntensityFree performs similarly or worse on Taxi, StackOverflow, Retweet, and Amazon, but better on Taobao, MIMIC-IV, and UKB (Table 17). On Retweet, the differences in log⁡p(t)\log p(t) mostly reflect how strongly a model exploits the discretized time grid. A continuous density can place narrow peaks on such a grid, and its log-likelihood then grows without bound. Scored by the probability of the recorded time bin instead (Appendix § 16), the gain from more knots vanishes on Retweet but persists on Amazon, whose time stamps lie on no grid (Figure 2).

Figure 2. Change in per-event test log p(t)p(t) and log p(d∣t)p(d \mid t) (nats) vs. spline knot count KK (relative to K=20K=20), for NeuralSplineTPP (intensity-head knots) and Spline/Spline Hawkes (kernel knots), on three datasets (all datasets in Tables 22 and 19). Left: the continuous time term log p(t)p(t) increases with KK on Retweet and Amazon. Middle: scored by the probability of the recorded time bin (Retweet: 1, Amazon: 0.01, MIMIC-IV: 1 min), it increases only on Amazon. Right: the type term log p(d∣t)p(d \mid t) does not improve with KK.

Table 17: IntensityFree with a GRU encoder (as in Shchur et al. [2020a]) vs. the Transformer encoder used in the main tables. Per-event test log-likelihood, mean ±\pm std over 5 seeds (higher is better). The Transformer is clearly better on Amazon, the GRU on MIMIC-IV and UKB and slightly on Taobao, and the two are close on Taxi, StackOverflow, and Retweet.

DatasetEncoderlog⁡p(t,d)\log p(t,d)log⁡p(t)\log p(t)log⁡p(d∣t)\log p(d\mid t)
TaxiTransformer0.44±.000.44\pm.000.73±.000.73\pm.00−0.29±.00-0.29\pm.00
GRU0.44±.000.44\pm.000.73±.000.73\pm.00−0.29±.00-0.29\pm.00
StackOverflowTransformer−2.33±.02-2.33\pm.02−0.73±.01-0.73\pm.01−1.60±.01-1.60\pm.01
GRU−2.31±.01-2.31\pm.01−0.70±.01-0.70\pm.01−1.61±.00-1.61\pm.00
RetweetTransformer−4.45±.26-4.45\pm.26−3.66±.26-3.66\pm.26−0.79±.01-0.79\pm.01
GRU−4.49±.27-4.49\pm.27−3.71±.27-3.71\pm.27−0.78±.02-0.78\pm.02
AmazonTransformer0.37±.040.37\pm.042.41±.012.41\pm.01−2.04±.03-2.04\pm.03
GRU−0.08±.13-0.08\pm.132.10±.132.10\pm.13−2.18±.06-2.18\pm.06
TaobaoTransformer1.30±.011.30\pm.012.73±.002.73\pm.00−1.42±.01-1.42\pm.01
GRU1.33±.011.33\pm.012.73±.012.73\pm.01−1.40±.00-1.40\pm.00
MIMIC-IVTransformer1.45±.031.45\pm.034.06±.024.06\pm.02−2.61±.01-2.61\pm.01
GRU1.69±.031.69\pm.034.22±.014.22\pm.01−2.53±.02-2.53\pm.02
UKBTransformer−12.87±.11-12.87\pm.11−7.42±.07-7.42\pm.07−5.44±.04-5.44\pm.04
GRU−12.34±.04-12.34\pm.04−7.19±.02-7.19\pm.02−5.16±.02-5.16\pm.02

Table 17.

The flexibility gain is a sharper time density. Decomposing each model’s log-likelihood into a time term log⁡p(t)\log p(t) and a time-conditional type term log⁡p(d∣t)\log p(d \mid t) reveals where the gains from flexibility arise. We vary the number of spline knots KK from 10 to 400, both in the intensity head of NeuralSplineTPP and in the kernels of Spline/Spline Hawkes (Figure 2, Table 22). For both models, log⁡p(t)\log p(t) increases by 0.9–1.6 nats on Retweet and Amazon and remains essentially unchanged on the other datasets, whereas log⁡p(d∣t)\log p(d \mid t) does not improve on any dataset. On Retweet, Amazon, and Taobao, the spread between history-dependent models is several times larger in log⁡p(t)\log p(t) than in log⁡p(d∣t)\log p(d \mid t) (Table 2). The exception is Taxi, whose event types strictly alternate between pick-up and drop-off events. Here, the type-conditional term separates the models that can represent inhibition (inhibitive Hawkes and neural TPPs) from the excitatory Hawkes processes. On these benchmarks, the flexibility that separates modern neural TPPs from classical Hawkes processes thus mostly sharpens when the next event occurs, not which type it is.

KKTaxiStackOverflowRetweetAmazonTaobaoMIMIC-IV
NeuralSplineTPP
100.72/−0.210.72/-0.21−0.73/−1.58-0.73/-1.58−3.32/−0.78-3.32/-0.781.63/−1.901.63/-1.902.66/−1.432.66/-1.433.99/−2.493.99/-2.49
200.73/−0.220.73/-0.22−0.72/−1.59-0.72/-1.59−3.32/−0.78-3.32/-0.781.41/−1.891.41/-1.892.70/−1.452.70/-1.453.99/−2.613.99/-2.61
500.71/−0.230.71/-0.23−0.73/−1.64-0.73/-1.64−3.13/−0.79-3.13/-0.792.22/−1.922.22/-1.922.72/−1.472.72/-1.473.97/−2.823.97/-2.82
1000.70/−0.240.70/-0.24−0.74/−1.69-0.74/-1.69−2.87/−0.78-2.87/-0.782.43/−1.902.43/-1.902.73/−1.492.73/-1.493.97/−2.893.97/-2.89
2000.67/−0.260.67/-0.26−0.75/−1.74-0.75/-1.74−2.44/−0.79-2.44/-0.792.53/−1.912.53/-1.912.72/−1.542.72/-1.543.95/−3.203.95/-3.20
4000.67/−0.270.67/-0.27−0.76/−1.76-0.76/-1.76−1.85/−0.79-1.85/-0.792.56/−1.912.56/-1.912.71/−1.602.71/-1.603.98/−3.033.98/-3.03
Spline/Spline Hawkes
100.48/−0.710.48/-0.71−0.73/−1.60-0.73/-1.60−3.36/−0.78-3.36/-0.780.14/−1.980.14/-1.982.66/−1.402.66/-1.403.69/−2.603.69/-2.60
200.49/−0.700.49/-0.70−0.73/−1.60-0.73/-1.60−3.36/−0.78-3.36/-0.780.43/−1.990.43/-1.992.67/−1.412.67/-1.413.83/−2.613.83/-2.61
500.48/−0.700.48/-0.70−0.73/−1.62-0.73/-1.62−3.29/−0.78-3.29/-0.781.29/−2.041.29/-2.042.67/−1.432.67/-1.433.83/−2.623.83/-2.62
1000.48/−0.710.48/-0.71−0.74/−1.65-0.74/-1.65−3.14/−0.79-3.14/-0.791.66/−2.031.66/-2.032.67/−1.462.67/-1.463.82/−2.643.82/-2.64
2000.49/−0.720.49/-0.72−0.76/−1.68-0.76/-1.68−2.84/−0.79-2.84/-0.791.70/−2.041.70/-2.042.66/−1.512.66/-1.513.81/−2.653.81/-2.65
4000.49/−0.740.49/-0.74−0.76/−1.71-0.76/-1.71−2.37/−0.80-2.37/-0.801.73/−2.041.73/-2.042.64/−1.552.64/-1.553.79/−2.683.79/-2.68

Table 22. Knot-count sweep on all datasets: log⁡p(t)\log p(t) / log⁡p(d∣t)\log p(d \mid t) (nats, mean over 5 seeds) for NeuralSplineTPP (head knots KK) and Spline/Spline Hawkes (kernel knots KK, baseline fixed at 100 knots). The sweep runs are separate from the reported runs (20,000 steps on the standard benchmarks, learning rates without per-model adjustment), so values differ from the main tables.

Benchmark datasets are close to saturation. A flexible baseline alone (Spline Poisson) performs considerably worse, whereas adding pairwise interactions (Inhib. Spline/Spline Hawkes) closes most of the gap to the neural TPPs. A goodness-of-fit test based on the time-rescaling theorem Brown et al., [4] (Table 25) shows that the flexible models are the best calibrated in time on four of the five datasets. As such, current benchmark datasets offer neural TPPs little opportunity to model interactions beyond pairwise ones.

Model (KS / ECE)TaxiStackOverflowRetweetAmazonTaobaoMIMIC-IV
Const./Exp Hawkes.151 / .274.033 / .023.060 / .008.267 / .045.063 / .022.315 / .040
Spline/Spline Hawkes.128 / .221.027 / .025.111 / .021.137 / .018.035 / .020.176 / .056
NeuralConstTPP.057 / .007.066 / .024.048 / .021.344 / .012.343 / .015.395 / .031
NeuralSplineTPP.040 / .008.009 / .027.083 / .009.025 / .015.037 / .016.021 / .024

Table 25. Calibration of four representative models with a closed-form cumulative intensity (seed 0). Each cell is the time-rescaling Kolmogorov–Smirnov distance sup⁡∣Femp(u)−u∣\sup\lvert F_{\mathrm{emp}}(u) - u\rvert of the probability integral transform ui=1−exp⁡(−Λ∗(ti−1,ti))u_i = 1 - \exp(-\Lambda^{*}(t_{i-1}, t_i)) (events with a zero inter-event time excluded) and the top-label expected calibration error of p(d∣t,H)p(d \mid t, \mathcal{H}) (15 bins); lower is better for both. Up to 10,000 test events per cell (MIMIC-IV: 5,000). For Spline/Spline Hawkes on StackOverflow, the evaluated checkpoint uses learning rate 0.1 instead of the adjusted rate of Table 8.

Large-Scale Medical Evaluation Tasks

Table 3 shows the results on the two medical tasks (full metrics in Table 15 and Table 16).

Table 15: Full MIMIC-IV results (D=117D = 117, including the OTHER class). log⁡p(t,d)=log⁡p(t)+log⁡p(d∣t)\log p(t,d) = \log p(t) + \log p(d\mid t); log⁡p(d)\log p(d), Acc and RMSE are computed by numerical integration on the first 500 test stays, all other metrics on the first 1,000. RMSE in days, multiplied by 100. {sup}‡Median predictor (see {xref}Table 14). {sup}§NeuralConstTPP uses a Softplus head on MIMIC-IV, as its exponential head diverged for individual seeds. Mean ± std over 5 seeds. Bold: best per column; underline: best classical (Poisson or Hawkes) model; ties at the displayed precision are all marked.

Modellog⁡p(t,d)\log p(t,d)log⁡p(t)\log p(t)log⁡p(d∣t)\log p(d\mid t)log⁡p(d)\log p(d)Acc∣t\mid t (%)Acc (%)RMSE×100\times100
Poisson processes
Homogeneous Poisson−1.18±.02-1.18\pm.022.60±.012.60\pm.01−3.78±.02-3.78\pm.02−3.81±.01-3.81\pm.0113.6±.013.6\pm.013.1±.013.1\pm.05.79±.015.79\pm.01
Spline Poisson−1.07±.01-1.07\pm.012.64±.002.64\pm.00−3.71±.01-3.71\pm.01−3.73±.01-3.73\pm.0113.6±.013.6\pm.013.1±.013.1\pm.05.75±.005.75\pm.00
Hawkes processes
Const./Exp Hawkes0.68±.000.68\pm.003.38±.003.38\pm.00−2.71±.00-2.71\pm.00−2.81±.00-2.81\pm.0030.9±.130.9\pm.130.0±.130.0\pm.15.48±.015.48\pm.01
Spline/Exp Hawkes0.68±.000.68\pm.003.39±.003.39\pm.00−2.70±.00-2.70\pm.00−2.81±.00-2.81\pm.0030.9±.130.9\pm.130.1±.130.1\pm.15.39±.015.39\pm.01
Const./Spline Hawkes1.16±.001.16\pm.003.78±.003.78\pm.00−2.62±.00-2.62\pm.00−2.79±.00-2.79\pm.0032.7±.132.7\pm.126.7±.126.7\pm.15.78±.015.78\pm.01
Spline/Spline Hawkes1.22±.001.22\pm.003.83±.003.83\pm.00−2.61±.00-2.61\pm.00−2.77±.00-2.77\pm.0032.9±.132.9\pm.125.9±.125.9\pm.15.96±.035.96\pm.03
Inhibitive Hawkes processes
Inhib. Const./Exp Hawkes1.19±.011.19\pm.013.87±.003.87\pm.00−2.67±.01-2.67\pm.01−2.76±.01-2.76\pm.0132.1±.232.1\pm.231.4±.231.4\pm.25.30±.015.30\pm.01
Inhib. Spline/Spline Hawkes1.14±.121.14\pm.123.83±.043.83\pm.04−2.68±.09-2.68\pm.09−2.73±.06-2.73\pm.0633.5±.333.5\pm.325.9±.425.9\pm.45.31±.045.31\pm.04
Neural TPPs
NeuralConstTPP{sup}§`0.43±.030.43\pm.033.08±.023.08\pm.02−2.65±.01-2.65\pm.01−2.64±.01-2.64\pm.0132.6±.332.6\pm.333.0±.333.0\pm.35.09±.025.09\pm.02
THP0.90±.010.90\pm.013.26±.013.26\pm.01−2.36±.01-2.36\pm.01−2.35±.01-2.35\pm.0137.6±.137.6\pm.138.2±.138.2\pm.14.97±.014.97\pm.01
SAHP0.92±.040.92\pm.043.66±.013.66\pm.01−2.74±.03-2.74\pm.03−2.76±.03-2.76\pm.0335.5±.435.5\pm.433.9±.133.9\pm.15.35±.025.35\pm.02
NHP0.62±.010.62\pm.013.60±.013.60\pm.01−2.98±.01-2.98\pm.01−2.92±.01-2.92\pm.0121.9±.221.9\pm.222.9±.222.9\pm.25.03±.025.03\pm.02
Flexible neural TPPs
IntensityFree1.45±.031.45\pm.034.06±.024.06\pm.02−2.61±.01-2.61\pm.01−2.60±.01-2.60\pm.0134.3±.334.3\pm.334.8±.434.8\pm.45.44±.045.44\pm.04{sup}‡`
NeuralSplineTPP1.67±.021.67\pm.024.05±.024.05\pm.02−2.39±.01-2.39\pm.01−2.50±.02-2.50\pm.0238.3±.138.3\pm.136.2±.236.2\pm.25.12±.025.12\pm.02

Table 15.

Table 16: Full UKB results (D=1,257D = 1{,}257, terminating). UKB models use the configuration of Appendix D.2, and IntensityFree uses its Transformer encoder. Inhibitive Hawkes processes were not evaluated on UKB (cost of numerical integration at D=1,257D = 1{,}257), and NHP did not converge. RMSE in days, divided by 100, using the predictive mean for all models. Mean ±\pm std over seeds. Bold: best per column; underline: best classical (Poisson or Hawkes) model; ties at the displayed precision are all marked.

Modellog⁡p(t,d)\log p(t,d)log⁡p(t)\log p(t)log⁡p(d∣t)\log p(d\mid t)log⁡p(d)\log p(d)Acc∣t\mid t (%)Acc (%)RMSE/100
Poisson processes
Homogeneous Poisson−13.67±.01-13.67\pm.01−8.08±.02-8.08\pm.02−5.59±.00-5.59\pm.00−5.59±.00-5.59\pm.004.23±.044.23\pm.044.26±.024.26\pm.0223.57±.2923.57\pm.29
Spline Poisson−12.87±.00-12.87\pm.00−7.51±.00-7.51\pm.00−5.36±.00-5.36\pm.00−5.40±.00-5.40\pm.004.56±.074.56\pm.074.37±.064.37\pm.0617.06±.1517.06\pm.15
Hawkes processes
Const./Exp Hawkes−12.51±.00-12.51\pm.00−7.36±.00-7.36\pm.00−5.15±.00-5.15\pm.00−5.03±.00-5.03\pm.006.45±.066.45\pm.066.91±.046.91\pm.0417.93±.0617.93\pm.06
Spline/Exp Hawkes−12.33±.00-12.33\pm.00−7.26±.00-7.26\pm.00−5.07±.00-5.07\pm.00−5.00±.00-5.00\pm.006.77±.106.77\pm.107.09±.047.09\pm.0415.26±.0115.26\pm.01
Const./Spline Hawkes−12.42±.00-12.42\pm.00−7.30±.00-7.30\pm.00−5.12±.00-5.12\pm.00−5.02±.00-5.02\pm.006.62±.056.62\pm.057.04±.047.04\pm.0417.48±.0717.48\pm.07
Spline/Spline Hawkes−12.28±.00-12.28\pm.00−7.23±.00-7.23\pm.00−5.05±.00-5.05\pm.00−4.99±.00-4.99\pm.007.00±.037.00\pm.037.22±.077.22\pm.0715.21±.0115.21\pm.01
Inhibitive Hawkes processes
Inhib. Const./Exp Hawkes–––––––
Inhib. Spline/Spline Hawkes–––––––
Neural TPPs
NeuralConstTPP−12.36±.01-12.36\pm.01−7.27±.00-7.27\pm.00−5.09±.01-5.09\pm.01−4.98±.01-4.98\pm.016.39±.056.39\pm.057.24±.057.24\pm.0514.68±.0314.68\pm.03
THP−12.80±.05-12.80\pm.05−7.39±.03-7.39\pm.03−5.41±.02-5.41\pm.02−5.40±.02-5.40\pm.024.75±.234.75\pm.234.69±.264.69\pm.2616.13±.1516.13\pm.15
SAHP−13.39±.05-13.39\pm.05−7.64±.02-7.64\pm.02−5.76±.03-5.76\pm.03−5.73±.03-5.73\pm.033.67±.233.67\pm.233.82±.323.82\pm.3218.49±.4318.49\pm.43
NHP–––––––
Flexible neural TPPs
IntensityFree−12.87±.11-12.87\pm.11−7.42±.07-7.42\pm.07−5.44±.04-5.44\pm.04−5.42±.05-5.42\pm.054.54±.194.54\pm.194.64±.224.64\pm.2217.24±.7817.24\pm.78
NeuralSplineTPP−12.21±.01-12.21\pm.01−7.19±.00-7.19\pm.00−5.02±.01-5.02\pm.01−4.96±.01-4.96\pm.017.07±.077.07\pm.077.22±.077.22\pm.0714.69±.0214.69\pm.02

Table 16.

Disease histories are well explained by additive interactions. Recent work has shown that transformer models capture population-scale disease trajectories across more than a thousand disease types [32]. Our UKB task is a simplified variant of this setting, excluding the sex and lifestyle covariates, evaluated on next-event likelihood and accuracy. On this task, flexible Hawkes processes with purely excitatory, additive interactions come close to the best neural model: Spline/Spline Hawkes trails NeuralSplineTPP by only 0.07 nats in joint log-likelihood and 0.08 percentage points in Acc∣t\mid t (Table 16), and outperforms other neural TPPs. This suggests that pairwise interactions capture most of the next-event structure in first-occurrence disease histories, making flexible Hawkes processes a strong and interpretable model choice. It does not speak to other capabilities of large generative models of health trajectories, such as long-horizon generation, or multi-modal extensions. We did not evaluate inhibitive Hawkes processes on UKB. However, the flexible neural models can represent inhibitive and non-additive effects, and they do not improve substantially on the excitatory Hawkes process.

MIMIC-IV: where do the gains come from?

On MIMIC-IV, NeuralSplineTPP leads the best Hawkes variant (Spline/Spline Hawkes) by 0.45 nats in joint log-likelihood, with 0.22 nats in the time term and 0.23 nats in the type term, and has a higher time-conditional accuracy (38.3% vs. 33.5% for the most type-accurate Hawkes variant, Inhib. Spline/Spline Hawkes, Table 15). THP and NeuralSplineTPP reach the best type terms (−2.36 and −2.39). IntensityFree, whose type distribution is independent of event time, performs at the level of the Hawkes processes (−2.61, and −2.53 with a GRU encoder, Table 17). The advantage of neural TPPs is concentrated on events recorded within the same minute as the preceding event, which make up 41.6% of all events. On these events, most neural TPPs predict the event type more accurately than the Hawkes processes. On other events, NeuralSplineTPP and Spline/Spline Hawkes differ by only 0.02 nats in log p(d∣t)p(d \mid t), and Spline/Spline Hawkes has the higher log p(t)p(t) (Appendix § 17). NeuralSplineTPP places its knots in the time elapsed since the last event, whereas the Hawkes spline kernels are anchored at each trigger event, which may favor the neural model on very short-range structure.

ModelUKBMIMIC-IV
H. Poisson−8.08/−5.592.60/−3.78
S. Poisson−7.51/−5.362.64/−3.71
C/E Hawkes−7.36/−5.153.38/−2.71
S/E Hawkes−7.26/−5.073.39/−2.70
C/S Hawkes−7.30/−5.123.78/−2.62
S/S Hawkes−7.23/−5.053.83/−2.61
I. C/E Hawkes–3.87/−2.67
I. S/S Hawkes–3.83/−2.68
N. Const−7.27/−5.093.08/−2.65
THP−7.39/−5.413.26/−2.36
SAHP−7.64/−5.763.66/−2.74
NHP–3.60/−2.98
IntensityFree−7.42/−5.444.06/−2.61
N. Spline−7.19/−5.024.05/−2.39

Table 3. Large-scale medical results (nats; cell = log p(t)p(t)/ log p(d∣t)p(d \mid t); row labels, bold and underline as in Table 2). On UKB, inhibitive Hawkes were not evaluated and NHP did not converge. Joint, ±std and full metrics in Appendix § 15.

Discussion

Our results suggest that prior benchmarks evaluated Hawkes processes in an overly restricted form. With flexible baseline intensities, learned kernels, and inhibition, Hawkes processes match or come close to the performance of neural TPPs on four of the five standard datasets. This indicates that these benchmarks are dominated by pairwise interactions. Amazon is the only standard benchmark on which the flexible neural models retain a clear advantage, mainly in the time term. On MIMIC-IV, the flexible neural TPPs retain an advantage in both terms, concentrated on events recorded in the same minute.

Neural TPPs such as THP [46], SAHP [42], NHP [21], RMTPP [6], and AttNHP [41] differ in their encoders but, like classical Hawkes processes, lack flexibility in the intensity parametrization. Splines offer a practical sweet spot for both families, as they are flexible yet allow exact sampling without thinning.

Implications for practitioners. Our experiments indicate that most of the datasets we consider are well described by pairwise interactions, and we expect this observation to extend to other TPP datasets [16]. Consequently, a flexible, well-tuned Hawkes process is a sensible starting point; the additional cost and complexity of a neural TPP are justified only for data with structure beyond pairwise interactions.

Implications for benchmarking and future developments. To assess neural TPPs, future benchmarks should use datasets with long-range dependencies and structure beyond pairwise interactions, such as electronic health records [37, 11], including rich event attributes beyond categorical types, such as lab measurements, genomic data, images, and clinical reports. Handling datasets requires care in how event times are recorded. On datasets with discretized timestamps, such as Retweet, the continuous-time likelihood is unbounded. Such data therefore require either dequantization or a discrete-time model. A related problem arises with simultaneous events, to which a continuous-time TPP assigns zero probability. Handling them requires discretized time, set-valued marks that represent all events at a time point jointly, or an explicit tie-breaking rule. More complex neural TPP methods [5, 13], especially those requiring substantial tuning or compute, should also be compared against simple but flexible alternatives like NeuralSplineTPP.

Limitations. (i) Scope. Our conclusions apply to Poisson and Hawkes processes, Transformer-based intensity models, NHP, and IntensityFree. They do not necessarily extend to Cox-type models or to flow-, diffusion-, or state-space-based TPPs, which we did not evaluate. (ii) Task. We evaluate autoregressive next-event prediction only. We do not assess long-horizon simulation or use of the models for planning and decision making, and calibration is checked only for four models with a closed-form cumulative intensity (Table 25). (iii) Datasets. Our conclusions rest on seven datasets, and other domains may favor neural models more prominently. Our aim is a fair comparison of TPP families, not state-of-the-art forecasting models for electronic health records. Both medical tasks are based on access-controlled data, and due to privacy regulations we can neither release any data derived from them nor the trained models. We use MIMIC-IV as a testbed for the modeling capabilities of TPPs, not to build a clinically useful forecasting model, and therefore retain only a subset of the recorded event types (Section 3.2). For use in continuous-time TPPs, we dequantize simultaneous events. On UKB, the sex and lifestyle covariates of the original pipeline [32] are not included for simplicity.

Conclusion

We have shown that the reported gap between classical Hawkes processes and neural TPPs is largely an artifact of under-specified models. Most of our datasets show little evidence of structure beyond pairwise interactions, highlighting the need for benchmarks that contain such structure. With flexible baseline intensities, flexible kernels, and inhibition, Hawkes processes match or come close to the performance of neural TPPs on four of the five standard benchmarks and remain competitive on UK Biobank disease histories. Neural TPPs are likewise limited by the flexibility of their intensity parametrization. We release TorchPP, a modular PyTorch framework with unified implementations of classical and neural TPPs, including support for terminating and right-censored sequences.

Acknowledgments and Disclosure of Funding

This work was supported by the European Research Council (ERC) under the European Union’s Horizon Europe research and innovation programme (project TAIPO, grant agreement No. 101088594). Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council Executive Agency. Neither the European Union nor the granting authority can be held responsible for them. This work was also funded by the Priority Program Translational Oncology of the Deutsche Krebshilfe (grant 70115167). Computations were performed on the GPU cluster of the DKFZ and on the UK Biobank Research Analysis Platform. This research has been conducted using the UK Biobank Resource under Application Number 81358. MIMIC-IV was accessed through PhysioNet under its credentialed-access data use agreement.

Implementation Overview

TORCHPP is organized around a single TemporalPointProcess abstraction that exposes intensity evaluation, cumulative-intensity computation, likelihood evaluation, density computation, distribution queries, and event sampling. Concrete models only need to provide the conditional intensity λt∗(t∣Ht)\lambda^*_{t}(t \mid\mathcal{H}_t) and, when available, a closed-form cumulative intensity. This keeps the training and evaluation code shared across Poisson, Hawkes, and neural families.

Core Abstractions

The base class separates three concerns. First, model-specific code computes the per-type intensity at an arbitrary query time. Second, the library provides shared likelihood and sampling utilities on top of that interface. Third, cumulative intensities can either be supplied analytically or approximated numerically through a common integration path. In practice, this means that simple models such as Poisson or exponential-kernel Hawkes processes use exact cumulative intensity computations, whereas inhibitive and fully neural models can still be trained and sampled through the same API with numerical quadrature.

The Hawkes family is implemented compositionally. A baseline module μd(t)\mu_d(t) and a kernel module ϕd′→d(τ)\phi_{d'\to d}(\tau) are instantiated independently and combined as

λd∗(t)=μd(t)+∑j:tj<tϕdj→d(t−tj),(3)\lambda_d^*(t) = \mu_d(t) + \sum_{j:t_j<t} \phi_{d_j\to d}(t-t_j), \tag*{(3)}

or, for inhibitive variants, after an outer positivity map gg,

λd∗(t)=g(μd(t)+∑j:tj<tϕdj→d(t−tj)).(4)\lambda_d^*(t) = g\left(\mu_d(t) + \sum_{j:t_j<t} \phi_{d_j\to d}(t-t_j)\right). \tag*{(4)}

This makes constant, linear, and spline baselines interchangeable with exponential, spline, and neural kernels without duplicating likelihood code.

Neural-network-based models use the same pattern at a different abstraction level. The event history is encoded once into a context representation hℓh_\ell per event at time tℓt_\ell, and a lightweight head maps (hℓ,t−tℓ)(h_\ell, t-t_\ell) to model parameters or directly to the intensity. Analytical heads such as the constant and spline heads provide closed-form cumulative intensities; fully neural heads fall back to the numerical path. Methods that integrate over the intensity function use a per-batch cache of the encoder output. Integration then requires only repeated evaluations of the lightweight head and a single forward pass through the encoder, which is substantially cheaper.

Support for Termination and Right-Censoring

Two implementation details are central for the medical data. First, terminating processes, as in the UK Biobank task, are handled explicitly: once an event type has occurred for a patient, the corresponding intensity is masked out and its cumulative intensity is truncated at the event time. This allows the same likelihood code to handle repeated-event and first-occurrence settings.

Second, right-censoring is supported. If a sequence is observed up to a sequence-specific censoring time C≥tNC \ge t_N, the survival term is integrated up to CC. We use this for MIMIC-IV, where the observation window ends at ICU discharge (for stays that are not truncated). For UKB, we set C=tNC = t_N. Previous neural TPP libraries assume that the censoring time coincides with the last event.

Sampling, Integration, and Reproducibility

TorchPP uses cumulative-intensity inversion as its default sampler, which is exact whenever the cumulative intensity is available in closed form. Given u∼Exp⁡(1)u \sim\operatorname{Exp}(1), the next event time is obtained by solving Λ∗(tℓ,t)=u\Lambda^*(t_\ell,t) = u, followed by categorical sampling over event types using the per-type intensities at the sampled time. Because the same cumulative-intensity code is also used in training, analytical and numerical models share the same execution path and the same validation checks. Ogata thinning remains available for comparison. For numerical integration, trapezoidal, Monte Carlo, and Monte Carlo trapezoidal rules are available.

At the systems level, all experiments are configured by composable YAML fragments with dotted command-line overrides. Training uses bucketed batches and dynamic batch sizing to reduce padding overhead and keep the memory footprint constant. The same model code is therefore used in the analytical derivations, the controlled benchmark runs, and the UKB experiments discussed in the main text.

Comparison to Previous Frameworks

EasyTPP is the closest public comparison point for the benchmark experiments, so we use it to highlight the design choices that materially affect evaluation.

Likelihood definition. EasyTPP evaluates an autoregressive likelihood that conditions on the first event,

log⁡p({(ti,di)}i=2N∣(t1,d1))=∑i=2Nlog⁡p(ti,di∣Hti),(5)\log p(\{(t_i,d_i)\}_{i=2}^{N} \mid(t_1,d_1)) = \sum_{i=2}^{N} \log p(t_i,d_i \mid\mathcal{H}_{t_i}), \tag*{(5)}

whereas TorchPP optimizes the full sequence likelihood

log⁡p(HT)=∑i=1Nlog⁡λdi∗(ti)−∫0Tλ∗(s) ds,(6)\log p(\mathcal{H}_T) = \sum_{i=1}^{N} \log\lambda_{d_i}^{*}(t_i) - \int_{0}^{T} \lambda^{*}(s)\,ds, \tag*{(6)}

which additionally contains the first-event term log⁡p(t1,d1)\log p(t_1,d_1) (including the survival on [0,t1][0,t_1]) and the survival term −∫tNTλ∗(s) ds-\int_{t_N}^{T}\lambda^{*}(s)\,ds after the last observed event. All TorchPP models are trained and selected on the full likelihood. All reported test log-likelihoods and accuracies follow Equation (5), averaged per predicted event (i=2,…,Ni=2,\ldots,N), for comparability with EasyTPP.

Support for terminating and censored sequences. EasyTPP assumes recurrent events on a shared observation window. TorchPP instead supports type-specific termination and per-sequence censoring times. In the terminating setting, each type-specific cumulative intensity is truncated at its first occurrence time; in the censored setting, the survival term is integrated up to the sequence-specific observation end.

Cumulative intensity evaluation. TorchPP switches between analytical and configurable numerical cumulative intensities (Monte Carlo, trapezoidal, and Monte Carlo trapezoidal) depending on the model. For neural models, encoder outputs are cached per batch and only the lightweight head is re-evaluated at quadrature points, which substantially lowers the cost of accurate numerical integration.

Improvements over the EasyTPP Benchmark Results

Xue et al. [ [40] used a fixed thinning over-sampling rate and a fixed search window for the proposal bound over all datasets (according to their repository). When the bound is too low, predicted event times are biased and RMSE is inflated, which we observed when we reran the tests with new configurations and obtained lower RMSE values than previously reported (see Appendix A.5). Our experiments also used a deterministic per-interval quadrature (Appendix D.4) and a longer training schedule with warmup and cosine learning-rate decay.

Validation of Implementation Correctness

To check the correctness of our implementations, we re-implemented the THP and SAHP architectures in TorchPP, matching their encoder dimensions, attention structure, activation pattern, and Softplus-wrapped intensity heads. Evaluation on the EasyTPP benchmark datasets additionally reports the same conditional likelihood convention used by EasyTPP, while TorchPP training remains on the full likelihood.

Table 4 shows that THP agrees closely across frameworks on accuracy, RMSE, and log-likelihood. SAHP matches in accuracy but exhibits larger likelihood discrepancies, especially on StackOverflow, which is consistent with the training-instability behavior observed in the EasyTPP implementation. In these runs, we integrate with 100 points during training, which stabilizes SAHP. The lower TorchPP RMSE values are expected once biased thinning is removed.

Table 4: Code equivalence validation: TorchPP re-implementations vs. EasyTPP framework on four benchmark datasets, with reference numbers from [40]. LL: joint log-likelihood (nats/event), Acc⁡∣t\operatorname{Acc}\lvert t: accuracy given the true next event time, as used by EasyTPP (as a fraction), RMSE: next-event time RMSE. TorchPP values are means over 5 seeds. EasyTPP (rerun) values are single-framework runs we performed locally. EasyTPP (paper) values are taken from the original EasyTPP benchmark; the reported type error rates are converted to accuracies. The TorchPP values use the EasyTPP-matched training configuration of Appendix § 11.§ 1 (20,000 steps, 100 integration points) and therefore differ from Appendix § 15. EasyTPP’s reported Hawkes baseline is analyzed separately in Appendix § 8.§ 6.

DatasetModelTorchPP LLTorchPP Acc⁡∣t\operatorname{Acc}\lvert tTorchPP RMSEEasyTPP (rerun) LLEasyTPP (rerun) Acc⁡∣t\operatorname{Acc}\lvert tEasyTPP (rerun) RMSEEasyTPP (paper) Acc⁡∣t\operatorname{Acc}\lvert tEasyTPP (paper) RMSE
TaxiSAHP0.250.9180.290−0.390.8970.3130.9030.372
THP0.250.9130.2850.240.9090.3050.9130.370
StackOverflowSAHP−2.360.4621.040−8.230.4251.3070.4391.375
THP−2.400.4521.029−2.540.4431.1300.4501.374
RetweetSAHP−4.140.59615.95−5.140.54019.690.58422.40
THP−4.170.59616.02−4.160.59717.570.58522.01
AmazonSAHP−2.150.3560.334−2.530.3390.3970.3230.619
THP−2.310.3570.331−2.220.3490.3740.3390.621

Table 4.

The close agreement suggests that the TorchPP re-implementations are faithful. Where metrics diverge, the remaining gap is explained by identified framework differences and by the sampling and integration issues in EasyTPP (see the difference between our EasyTPP re-runs and the paper values), rather than by a mismatch in model equations.

Reproducing EasyTPP’s Hawkes Baseline

EasyTPP [40] reports a multivariate Hawkes (MHP) baseline whose next-type accuracy is far below ours on four of the five datasets and above our non-inhibitive Hawkes processes on Taxi (90.5%). According to their Appendix B, this baseline was fitted externally with the tick library [1] using an exponential kernel with fixed parameters: “the base intensity is set to 0.2 for all dimensions while the decay matrix is set to be a unit matrix plus a small noise”. The code that produced these numbers is not part of the EasyTPP repository, which contains neither a Hawkes model nor tick-fitting code, so the numbers cannot be re-derived from the released code. We therefore re-fitted the same model with tick (version 0.8.0.2) following this recipe on the native EasyTPP time units, and evaluated Acc⁡∣t\operatorname{Acc}|t as arg⁡max⁡dλλd∗(ti)\arg\max_{d_\lambda}\lambda_{d}^{*}(t_i) at the true next-event time with a causal history, the convention of EasyTPP’s Appendix § 9.§ 2 and of our own evaluation. As tick’s likelihood-based estimator fails in this release, we additionally fitted the identical model by positivity-constrained maximum likelihood.

Table 5 shows the results. Our tick fits match or exceed EasyTPP’s reported MHP accuracy on StackOverflow, Retweet, Amazon, and Taobao (the one exception is the fit with the base intensity fixed to 0.2 on Retweet, far from that dataset’s natural rate), where the reported values lie 6–12 points below the majority-class rate. No properly fitted model should fall below this value, since always predicting the most frequent type attains it. On Taxi, no faithful excitation-only fit exceeds ≈ 85%, in line with our learned Const./Exp Hawkes. The 90% regime on Taxi is reached only by our inhibitive Hawkes processes and the neural models, since a Hawkes process with non-negative interactions cannot suppress a type that has just occurred, which the Taxi data appears to require. The same pipeline reproduces EasyTPP’s neural accuracies closely (Table 4). We conclude that the discrepancy is specific to EasyTPP’s MHP baseline and that we cannot compare to it in good faith. We therefore do not include its numbers in our main tables.

Table 5: Reproducing EasyTPP’s Hawkes (MHP) baseline. Time-conditional next-type accuracy Acc⁡∣t\operatorname{Acc}\lvert t (%). The tick fits, the majority rate and EasyTPP use the full test sets, our models the first 1,000 test sequences (§ 3.3). Majority: always predicting the most frequent type. tick LS: tick HawkesExpKern with fixed decay 1.0 and its default least-squares fit; MLE: the same exponential Hawkes fitted by positivity-constrained maximum likelihood; MLE, μ\mu=0.2: additionally fixing the base intensity to 0.2, the literal recipe of EasyTPP’s Appendix § 9. Ours: our learned Const./Exp Hawkes and the best of our (inhibitive) Hawkes variants, from Table 11.

DatasetMajoritytick LSMLEMLE, μ\mu=0.2EasyTPP (reported)Ours: C/E / best Hawkes
Taxi44.379.684.284.690.585.0 / 92.0
StackOverflow42.544.543.644.235.046.2 / 46.7
Retweet54.053.854.142.244.357.5 / 58.6
Amazon30.029.928.932.824.135.0 / 39.1
Taobao43.654.153.754.531.961.0 / 61.5

Table 5.

Table 11: Time-conditional top-1 type accuracy Acc⁡∣t\operatorname{Acc}|t (%, higher is better): arg⁡max⁡dp(d∣t,H)\arg\max_d p(d \mid t,\mathcal{H}) at the true next-event time, the next-type metric reported by EasyTPP [40]. Mean ±\pm std over 5 seeds. Bold: best per dataset; underline: best classical (Poisson or Hawkes) model; ties at the displayed precision are all marked.

ModelTaxiStackOverflowRetweetAmazonTaobao
Poisson processes
Homogeneous Poisson44.29±.0044.29\pm.0042.50±.0042.50\pm.0053.92±.0053.92\pm.0029.87±.0029.87\pm.0043.57±.0043.57\pm.00
Spline Poisson44.26±.1044.26\pm.1042.50±.0042.50\pm.0054.66±.0654.66\pm.0629.87±.0029.87\pm.0043.57±.0043.57\pm.00
Hawkes processes
Const./Exp Hawkes85.00±.0885.00\pm.0846.16±.0446.16\pm.0457.49±.1657.49\pm.1635.03±2.5035.03\pm2.5060.96±.0560.96\pm.05
Spline/Exp Hawkes84.74±.1284.74\pm.1246.16±.0846.16\pm.0858.65±.1058.65\pm.1032.48±.0132.48\pm.0160.91±.0660.91\pm.06
Const./Spline Hawkes83.28±.0983.28\pm.0946.72±.0446.72\pm.0457.24±.2057.24\pm.2038.04±.0838.04\pm.0861.55±.0561.55\pm.05
Spline/Spline Hawkes83.50±.0283.50\pm.0246.42±.1746.42\pm.1756.55±.1256.55\pm.1235.87±.0335.87\pm.0361.54±.0661.54\pm.06
Inhibitive Hawkes processes
Inhib. Const./Exp Hawkes90.95±.0490.95\pm.0446.41±.0346.41\pm.0357.75±.2657.75\pm.2632.61±.9632.61\pm.9661.06±.0261.06\pm.02
Inhib. Spline/Spline Hawkes92.00±.0292.00\pm.0246.07±.3346.07\pm.3357.65±.0857.65\pm.0839.15±.2139.15\pm.2161.23±.1161.23\pm.11
Neural TPPs
NeuralConstTPP91.44±.0691.44\pm.0645.24±.1145.24\pm.1158.55±1.2058.55\pm1.2035.62±.0635.62\pm.0660.23±.1660.23\pm.16
THP91.47±.0691.47\pm.0645.22±.1645.22\pm.1659.28±.3059.28\pm.3035.70±.0635.70\pm.0660.33±.1760.33\pm.17
SAHP92.35±.2792.35\pm.2746.79±.0846.79\pm.0858.24±.3058.24\pm.3032.29±.7532.29\pm.7560.34±.5760.34\pm.57
NHP92.26±.3292.26\pm.3246.76±.1746.76\pm.1756.20±2.3656.20\pm2.3633.30±1.0533.30\pm1.0560.91±.1260.91\pm.12
Flexible neural TPPs
IntensityFree91.49±.0791.49\pm.0745.23±.1645.23\pm.1658.69±.9458.69\pm.9434.87±.6634.87\pm.6660.14±.2760.14\pm.27
NeuralSplineTPP92.66±.0592.66\pm.0546.61±.1646.61\pm.1658.70±.5758.70\pm.5739.30±.4839.30\pm.4860.42±.3060.42\pm.30

Table 11.

Mathematical Details on Temporal Point Processes

This appendix provides a self-contained derivation of the key quantities used throughout the paper: the conditional density, cumulative intensity, survival function, log-likelihood, marginal distributions, and sampling algorithms. Figure 3 summarizes how these objects relate to one another.

Relationships between the key quantities of a multivariate temporal point process

Figure 3. Relationships between the key quantities of a multivariate temporal point process. The top row constructs the survival objects from the per-type conditional intensities λd∗(t)\lambda_d^*(t): summation across types yields the ground intensity λ∗(t)\lambda^*(t), integration over time yields the cumulative intensity Λ∗(ti−1,t)\Lambda^*(t_{i-1},t), and the survival S∗(τ)S^*(\tau) and CDF F∗(τ)F^*(\tau) of the next-event time follow from Λ∗\Lambda^* by exp⁡(−⋅)\exp(-\cdot) and 1−⋅1-\cdot. The middle row shows how the next-event densities are constructed at merge nodes (×\times, //): each density is formed by combining two top-row quantities, and the resulting expression is the label on the outgoing arrow. In addition, the time marginal is recovered as p∗(t)=∑dp∗(t,d)p^*(t)=\sum_d p^*(t,d), the type marginal as p∗(d)=∫ti−1∞p∗(t,d) dtp^*(d)=\int_{t_{i-1}}^\infty p^*(t,d)\,dt, and the time conditional as p∗(t∣d)=p∗(t,d)/p∗(d)p^*(t\mid d)=p^*(t,d)/p^*(d). Quantities shown in red (p∗(d)p^*(d) and p∗(t∣d)p^*(t\mid d)) are expensive to evaluate when starting from intensities: both require the unbounded integral defining p∗(d)p^*(d), which has no closed form except for trivial cases such as the homogeneous Poisson process and otherwise demands numerical quadrature with Λ∗\Lambda^* evaluated at every quadrature point. Stars denote conditioning on the history Hti−1\mathcal{H}_{t_{i-1}}; τ=t−ti−1\tau=t-t_{i-1}.

Conditional Intensity and Cumulative Intensity

A temporal point process on [0,T][0,T] is fully characterized by its conditional intensity function

λ∗(t)=lim⁡Δt→0Pr⁡(event in [t,t+Δt)∣Ht)Δt,(7)\lambda^*(t)=\lim_{\Delta t\to0}\frac{\Pr(\text{event in }[t,t+\Delta t)\mid\mathcal{H}_t)}{\Delta t}, \tag*{(7)}

where Ht={(ti,di):ti<t}\mathcal{H}_t=\{(t_i,d_i):t_i<t\} denotes the history of events before time tt. In the multivariate case with DD event types, the ground (total) intensity is λ∗(t)=∑d=1Dλd∗(t)\lambda^*(t)=\sum_{d=1}^{D}\lambda_d^*(t).

The cumulative intensity between times aa and bb is defined as

Λ∗(a,b)=∫abλ∗(s) ds=∑d=1D∫abλd∗(s) ds=∑d=1DΛd∗(a,b),(8)\Lambda^*(a,b)=\int_a^b\lambda^*(s)\,ds=\sum_{d=1}^{D}\int_a^b\lambda_d^*(s)\,ds=\sum_{d=1}^{D}\Lambda_d^*(a,b), \tag*{(8)}

which is the compensator of the counting process. Within one inter-event interval, where the history is fixed, it is the cumulative hazard of the next event.

Survival Function and the Inter-Event Distribution

Consider the ii-th inter-event interval. Let τ=t−ti−1\tau=t-t_{i-1} denote the time elapsed since the last event. Here and below, conditioning on Hti−1\mathcal{H}_{t_{i-1}} is shorthand for Ht\mathcal{H}_t with t∈(ti−1,ti]t\in(t_{i-1},t_i], i.e. the history including (ti−1,di−1)(t_{i-1}, d_{i-1}), which is fixed within the interval. The probability that no event of any type occurs in (ti−1,ti−1+τ](t_{i-1}, t_{i-1}+\tau] is given by the survival function:

S∗(τ)=Pr⁡(no event in (ti−1,ti−1+τ]∣Hti−1)=exp⁡(−Λ∗(ti−1,ti−1+τ))=exp⁡(−∫ti−1ti−1+τλ∗(s) ds).(9)S^{*}(\tau) = \Pr(\text{no event in }(t_{i-1},t_{i-1}+\tau] \mid\mathcal{H}_{t_{i-1}}) = \exp\left(-\Lambda^{*}(t_{i-1},t_{i-1}+\tau)\right) = \exp\left(-\int_{t_{i-1}}^{t_{i-1}+\tau}\lambda^{*}(s)\,\mathrm{d}s\right). \tag*{(9)}

This is the multivariate analogue of the standard survival function in survival analysis: the cumulative intensity acts as the cumulative hazard.

The density of the next event occurring at time tt with type dd, conditioned on history, is

p∗(t,d∣Hti−1)=λd∗(t)⋅S∗(t−ti−1)=λd∗(t)exp⁡(−Λ∗(ti−1,t)).(10)p^{*}(t,d \mid\mathcal{H}_{t_{i-1}}) = \lambda^{*}_{d}(t)\cdot S^{*}(t-t_{i-1}) = \lambda^{*}_{d}(t)\exp\left(-\Lambda^{*}(t_{i-1},t)\right). \tag*{(10)}

Intuition: the event must happen at time tt (rate λd∗(t)\lambda^{*}_{d}(t)) and nothing must have happened before tt (survival S∗S^{*}).

Deriving the Log-Likelihood

The log-likelihood of a complete sequence HT={(t1,d1),…,(tN,dN)}\mathcal{H}_{T}=\{(t_{1},d_{1}),\ldots,(t_{N},d_{N})\} observed on [0,T][0,T] factorizes over inter-event intervals. We sum the NN event contributions from Equation (10) together with the survival probability that no event occurs between tNt_{N} and the observation end TT:

log⁡p(HT)=∑i=1Nlog⁡p∗(ti,di∣Hti−1)+log⁡S∗(T−tN)=∑i=1N[log⁡λdi∗(ti)−Λ∗(ti−1,ti)]−Λ∗(tN,T)=∑i=1Nlog⁡λdi∗(ti)−∑i=1NΛ∗(ti−1,ti)−Λ∗(tN,T)=∑i=1Nlog⁡λdi∗(ti)−Λ∗(0,T)=∑i=1Nlog⁡λdi∗(ti)⏟event log-intensities−∫0Tλ∗(s) ds⏟cumulative intensity.(11)\begin{aligned} \log p(\mathcal{H}_{T}) &= \sum_{i=1}^{N}\log p^{*}(t_{i},d_{i}\mid\mathcal{H}_{t_{i-1}})+\log S^{*}(T-t_{N})\\ &= \sum_{i=1}^{N}\left[\log\lambda^{*}_{d_{i}}(t_{i})-\Lambda^{*}(t_{i-1},t_{i})\right]-\Lambda^{*}(t_{N},T)\\ &= \sum_{i=1}^{N}\log\lambda^{*}_{d_{i}}(t_{i})-\sum_{i=1}^{N}\Lambda^{*}(t_{i-1},t_{i})-\Lambda^{*}(t_{N},T)\\ &= \sum_{i=1}^{N}\log\lambda^{*}_{d_{i}}(t_{i})-\Lambda^{*}(0,T)\\ &= \underbrace{\sum_{i=1}^{N}\log\lambda^{*}_{d_{i}}(t_{i})}_{\text{event log-intensities}}-\underbrace{\int_{0}^{T}\lambda^{*}(s)\,\mathrm{d}s}_{\text{cumulative intensity}}. \tag*{(11)} \end{aligned}

The combination in the penultimate step uses Λ∗(0,T)=∑i=1NΛ∗(ti−1,ti)+Λ∗(tN,T)\Lambda^{*}(0,T)=\sum_{i=1}^{N}\Lambda^{*}(t_{i-1},t_{i})+\Lambda^{*}(t_{N},T) with t0≡0t_{0}\equiv0. This recovers the main-text log-likelihood expression.

The first term rewards high intensity at observed event times, while the second term penalizes high intensity everywhere, acting as a normalizer.

Decomposing the Joint Density

The joint conditional density from Equation (10) admits a natural decomposition into a time marginal and a type conditional.

Marginal time density p∗(t∣H)p^{*}(t\mid\mathcal{H}). Summing Equation (10) over all types dd:

p∗(t∣Hti−1)=∑d=1Dp∗(t,d∣Hti−1)=λ∗(t)exp⁡(−Λ∗(ti−1,t)).(12)p^{*}(t\mid\mathcal{H}_{t_{i-1}})=\sum_{d=1}^{D}p^{*}(t,d\mid\mathcal{H}_{t_{i-1}})=\lambda^{*}(t)\exp\left(-\Lambda^{*}(t_{i-1},t)\right). \tag*{(12)}

This is the density of the next-event time regardless of its type. It has the familiar form of hazard times survival.

Conditional type distribution p∗(d∣t,H)p^*(d \mid t, \mathcal{H}). Dividing Equation (10) by Equation (12):

p∗(d∣t,H)=p∗(t,d∣H)p∗(t∣H)=λd∗(t)λ∗(t)=λd∗(t)∑d′=1Dλd′∗(t).(13)p^*(d \mid t, \mathcal{H}) = \frac{p^*(t,d \mid\mathcal{H})}{p^*(t \mid\mathcal{H})} = \frac{\lambda_d^*(t)}{\lambda^*(t)} = \frac{\lambda_d^*(t)}{\sum_{d'=1}^{D}\lambda_{d'}^*(t)}. \tag*{(13)}

Given the next-event time tt, the type is sampled categorically with probabilities proportional to the per-type intensities. Note that the survival terms cancel exactly, so the type distribution depends only on the intensity ratios at time tt.

Marginal type distribution p∗(d∣H)p^*(d \mid\mathcal{H}). The probability that the next event is of type dd, regardless of when it occurs, is obtained by marginalizing out the time:

p∗(d∣Hti−1)=∫ti−1∞p∗(t,d∣Hti−1) dt=∫ti−1∞λd∗(t)exp⁡(−Λ∗(ti−1,t)) dt.(14)p^*(d \mid\mathcal{H}_{t_{i-1}}) = \int_{t_{i-1}}^{\infty} p^*(t,d \mid\mathcal{H}_{t_{i-1}})\,dt = \int_{t_{i-1}}^{\infty} \lambda_d^*(t)\exp\left(-\Lambda^*(t_{i-1},t)\right)\,dt. \tag*{(14)}

This integral generally has no closed form, except when the type distribution does not depend on tt (e.g. for time-constant intensities such as the homogeneous Poisson process or NeuralConstTPP, or for IntensityFree), and must otherwise be computed numerically, for example by quadrature over the next-event time density.

Sampling: Inversion Method

The time-rescaling theorem [4] states that if {ti}\{t_i\} are event times of a point process with intensity λ∗(t)\lambda^*(t), then the rescaled times Λ∗(0,ti)\Lambda^*(0,t_i) form a unit-rate Poisson process, that is, the increments ui=Λ∗(ti−1,ti)u_i = \Lambda^*(t_{i-1},t_i) are i.i.d. Exp⁡(1)\operatorname{Exp}(1). This requires Λ∗(ti−1,∞)=∞\Lambda^*(t_{i-1},\infty)=\infty, which holds for the closed-form models evaluated in Table 25, as their baseline and head splines are extrapolated beyond the last knot as a constant. To see this, note the relationship between CDF, survival, and cumulative intensity:

  1. The survival function is S∗(τ)=exp⁡(−Λ∗(ti−1,ti−1+τ))S^*(\tau)=\exp(-\Lambda^*(t_{i-1},t_{i-1}+\tau)).

  2. The CDF is F∗(τ)=1−S∗(τ)=1−exp⁡(−Λ∗(ti−1,ti−1+τ))F^*(\tau)=1-S^*(\tau)=1-\exp(-\Lambda^*(t_{i-1},t_{i-1}+\tau)).

  3. Substituting u=Λ∗(ti−1,ti−1+τ)u=\Lambda^*(t_{i-1},t_{i-1}+\tau) gives Pr⁡(ui≤u∣Hti−1)=1−e−u\Pr(u_i\le u\mid\mathcal{H}_{t_{i-1}})=1-e^{-u}, so ui∼Exp⁡(1)u_i\sim\operatorname{Exp}(1) independently of the history.

This provides an exact sampling algorithm via cumulative intensity inversion:

  1. Draw u∼Exp⁡(1)u\sim\operatorname{Exp}(1), equivalently u=−log⁡(1−v)u=-\log(1-v) with v∼Uniform⁡(0,1)v\sim\operatorname{Uniform}(0,1).

  2. Find t∗t^* such that Λ∗(ti−1,t∗)=u\Lambda^*(t_{i-1},t^*)=u, that is, invert the cumulative intensity.

  3. Sample the event type d∗∼Categorical⁡(λ1∗(t∗)/λ∗(t∗),…,λD∗(t∗)/λ∗(t∗))d^*\sim\operatorname{Categorical}(\lambda_1^*(t^*)/\lambda^*(t^*),\ldots,\lambda_D^*(t^*)/\lambda^*(t^*)).

  4. If t∗>Tt^*>T, no further events occur; otherwise record (t∗,d∗)(t^*,d^*) and repeat.

Step 2 requires solving a root-finding problem: find t∗t^* such that ∫ti−1t∗λ∗(s) ds=u\int_{t_{i-1}}^{t^*}\lambda^*(s)\,ds=u. When the cumulative intensity is available in closed form, this is solved with fast bracketed Newton iterations. For models without a closed-form cumulative intensity, numerical quadrature is used to evaluate Λ∗\Lambda^* within the root finder.

Advantages over thinning. Inversion does not require an upper bound on the intensity, so it is unbiased by construction whenever the cumulative intensity can be evaluated. It is also more efficient when the intensity varies by orders of magnitude over the observation window, as is common in medical event data with long quiescent periods. However, if no closed-form cumulative intensity is available, and an upper bound on the intensity can be estimated, Ogata’s thinning algorithm is more reliable, as it does not compound integration errors.

Sampling: Ogata’s Thinning Algorithm

Ogata’s thinning algorithm [24] is the classical approach for sampling from a TPP with intensity λ∗(t)\lambda^*(t):

  1. Starting from the last event time ti−1t_{i-1}, compute an upper bound λˉ≥sup⁡s>ti−1λ∗(s)\bar{\lambda} \ge\sup_{s>t_{i-1}} \lambda^*(s), or a local upper bound on some interval.

  1. Draw a candidate inter-event time Δt∼Exp⁡(λˉ)\Delta t \sim\operatorname{Exp}(\bar{\lambda}).

  1. Accept the candidate s=ti−1+Δts = t_{i-1} + \Delta t with probability λ∗(s)/λˉ\lambda^*(s)/\bar{\lambda}; otherwise reject it and repeat from ss.

  1. If accepted, sample the type as in the inversion method.

The overall acceptance rate is the average of λ∗(s)/λˉ\lambda^*(s)/\bar{\lambda} over the proposed points. When λˉ\bar{\lambda} is much larger than the typical intensity, for example due to rare high-intensity bursts, the algorithm becomes very inefficient, generating many rejected proposals. Adaptive thinning with piecewise upper bounds mitigates this but adds implementation complexity. Critically, if the upper bound is set too low, below the true supremum, the algorithm produces biased samples. This is especially a problem for neural TPPs, as even finding the upper bound over an interval can be challenging.

Model Details

Let tℓt_\ell denote the most recent event time before a query time tt, let τ=t−tℓ\tau= t - t_\ell, and let hℓh_\ell be the context embedding summarizing the history up to tℓt_\ell whenever a neural encoder is used.

Baseline-Only Models

The homogeneous Poisson model uses a constant per-type rate,

λd∗(t)=μd.(15)\lambda_d^*(t) = \mu_d. \tag*{(15)}

The inhomogeneous baseline-only models replace the constant rate by a learned time-dependent baseline μd(t)\mu_d(t), implemented in TorchPP as spline interpolation over a small set of knots. For spline Poisson models this yields

λd∗(t)=μd(t),(16)\lambda_d^*(t) = \mu_d(t), \tag*{(16)}

with analytical cumulative intensity given by the integral of the corresponding spline. The spline knot heights and the constant base rate are passed through a softplus to ensure non-negativity. Beyond the last knot, the baseline spline is extrapolated as a constant.

Additive Hawkes Models

All Hawkes variants share the additive structure

λd∗(t)=μd(t)+∑j:tj<tϕdj→d(t−tj).(17)\lambda_d^*(t) = \mu_d(t) + \sum_{j:t_j<t} \phi_{d_j\to d}(t-t_j). \tag*{(17)}

The models differ only in the chosen baseline μd\mu_d and kernel ϕd′→d\phi_{d'\to d}. For exponential-kernel Hawkes models,

ϕd′→d(τ)=αd′dβd′dexp⁡(−βd′dτ),τ>0,(18)\phi_{d'\to d}(\tau) = \alpha_{d'd}\beta_{d'd}\exp(-\beta_{d'd}\tau), \qquad\tau> 0, \tag*{(18)}

which gives a closed-form cumulative intensity. Spline-kernel variants keep the same additive decomposition but replace the exponential kernel by a linear spline over elapsed time. Its integral, and hence the cumulative intensity, is a monotone piecewise-quadratic function in closed form.

Inhibitive Hawkes Models

To model inhibition, TorchPP wraps the additive pre-activation in a positivity map gg, typically Softplus,

λd∗(t)=g(μd(t)+∑j:tj<tϕdj→d(t−tj)).(19)\lambda_d^*(t) = g\left(\mu_d(t) + \sum_{j:t_j<t} \phi_{d_j\to d}(t-t_j)\right). \tag*{(19)}

This allows negative intermediate contributions while preserving a non-negative intensity. Because the outer nonlinearity destroys the closed-form integral, these models use numerical quadrature for the cumulative intensities.

Transformer-Conditioned Neural TPPs

The simplest neural family predicts a history-dependent intensity that is constant between events,

λd∗(t)=exp⁡(wdThℓ+bd).(20)\lambda_d^*(t) = \exp\left(w_d^{\mathsf{T}} h_\ell+ b_d\right). \tag*{(20)}

which corresponds to the NeuralConstTPP model. The NeuralSplineTPP model predicts spline knot heights from hℓh_\ell and evaluates the intensity as a positive spline function of τ\tau, with analytical integration over each spline segment. The same Transformer backbone also supports THP- and SAHP-style heads, whose softplus intensities are integrated numerically.

Recurrent and Intensity-Free Neural TPPs

NHP [21] encodes the history with a continuous-time LSTM whose cell state decays between events, and maps the hidden state at time tt to per-type intensities with a scaled softplus. Its cumulative intensity is integrated numerically. IntensityFree [29] is a density-based model: instead of an intensity it parametrizes the next inter-event time directly as a mixture of 64 log-normal components, p(τ∣hℓ)=∑kwk(hℓ)LogNormal⁡(τ;μk(hℓ),σk(hℓ))p(\tau\mid h_\ell) = \sum_k w_k(h_\ell)\operatorname{LogNormal}(\tau; \mu_k(h_\ell), \sigma_k(h_\ell)), and the next type by a categorical distribution p(d∣hℓ)p(d \mid h_\ell) that does not depend on τ\tau. Both are available in closed form, so no numerical integration is needed. By default we use a Transformer encoder, and the GRU encoder of the original model is compared in Table 17.

Hyperparameter Tuning and Training Details

Hyperparameters were selected by staged grid searches implemented directly in the repository scripts. The EasyTPP benchmarks, UK Biobank, and MIMIC-IV use separate search pipelines because they differ substantially in sequence length, censoring, and the stability of numerical models. All three pipelines share the same overall structure described below; per-dataset deviations are listed in the corresponding subsections.

Scripts with the exact configurations for reproduction can be found in our repository.

Common training settings. All tuning runs use AdamW with linear warmup followed by cosine annealing to ηmin⁡=LR/100\eta_{\min} = \mathrm{LR}/100, gradient clipping at 1.0, and per-event normalized loss. Validation log-likelihood is computed every 250 optimization steps on the standard benchmarks (every 500 on UKB and MIMIC-IV), and early stopping with patience 4 (6 on UKB and MIMIC-IV) is applied on the validation joint log-likelihood. Bucketed dynamic batching is used throughout: a small-bucket preset (bucket_default_small) for analytical Hawkes and standard neural runs, and a smaller numerical preset (bucket_numerical_small) for the inhibitive Hawkes variants, which require numerical cumulative-intensity integration. On the standard benchmarks, spline baselines use Tfactor=1.25T_{\mathrm{factor}} = 1.25, and spline kernels use a knot-power schedule of 4.0 (denser knots near 0 to capture short-range excitation). The UKB and MIMIC-IV settings are given below.

Two-stage search. The grid is partitioned into two stages so that expensive models inherit architectural and time-scale choices from cheaper ones:

  • Stage 1: cheap analytical families, i.e. Poisson, Spline Poisson (sweeping the baseline knot count KbK_b), constant-baseline exponential Hawkes, constant-baseline spline-kernel Hawkes (sweeping kernel knots KkK_k and time-scale factor TfT_f), and NeuralConstTPP (sweeping Transformer depth LL, heads HH, and embedding dimension dd). Some pipelines also include variants that are not reported in this paper. Stage 1 selects the best KbK_b, (Kk,Tf)(K_k,T_f), and (L,H,d)(L,H,d) per dataset, which fixes the architectural backbone and the inter-event time scale for the rest of the search.

  • Stage 2: expressive analytical models and neural heads conditioned on the Stage 1 winners, i.e. spline-baseline exponential Hawkes, spline-baseline spline-kernel Hawkes, the inhibitive Hawkes variants with softplus clamping, and NeuralSplineTPP (with a small refinement grid for (Kk,Tf)(K_k,T_f) on top of the Stage 1 backbone).

The exact per-dataset grids are summarized in the subsections below.

EasyTPP Benchmarks

The benchmark search is orchestrated by scripts/tuning/tune_easytpp_combined_local.sh. Tuning runs use 20,000 optimization steps, learning rate 0.1 for Hawkes-family models and 0.01 for neural models, 500-step warmup, and the smaller bucket preset. Stage 1 sweeps Kb∈{5,10,15}K_b \in\{5, 10, 15\} for Spline Poisson; Kk∈{2,4,8}K_k \in\{2, 4, 8\} and Tf∈{5,10,15,20}T_f \in\{5, 10, 15, 20\} (multiples of the mean inter-event time) for the spline-kernel Hawkes; and the transformer backbone over L∈{2,4}L \in\{2, 4\}, H∈{2,4}H \in\{2, 4\}, d∈{32,64}d \in\{32, 64\}. Stage 2 reuses the Stage 1 winners and adds a refinement grid Kk∈{2,4,8}K_k \in\{2, 4, 8\}, Tf∈{5,10,15,20}T_f \in\{5, 10, 15, 20\} for NeuralSplineTPP. The configuration of the reported runs is given in Appendix D.4.

UK Biobank

The UKB search is implemented in scripts/tuning/tune_ukb_real_combined_local.sh. Because the vocabulary is much larger and sequences span decades, the defaults are more conservative: 50,000 optimization steps and learning rate 0.01 for Hawkes families and 0.001 for neural families. Stage 1 sweeps Kb∈{15,20}K_b \in\{15, 20\} for Spline Poisson, (Kk,Tf)∈{5,8,10}×{15,20}(K_k, T_f) \in\{5, 8, 10\} \times\{15, 20\} for the spline-kernel Hawkes, and the transformer backbone over L∈{8,12}L \in\{8, 12\}, H∈{8,16}H \in\{8, 16\}, d∈{64,128,256}d \in\{64, 128, 256\} (substantially wider than the EasyTPP grid, reflecting the 1,257-type vocabulary). Stage 2 reuses the Stage 1 backbone and sweeps NeuralSplineTPP over (Kk,Tf)∈{5,8,10}×{15,20}(K_k, T_f) \in\{5, 8, 10\} \times\{15, 20\}. Inhibitive Hawkes processes were not evaluated on UKB because of the cost of numerical integration at D=1,257D = 1{,}257. The reported UKB runs are defined in scripts/tuning/generate_replica_commands_ukb.sh. They use spline time-scale factors Tfactor=1.0T_{\mathrm{factor}} = 1.0, a kernel knot power of 4.5, 100 integration points for the numerically integrated neural models, and validation on 2,000 sequences. THP and SAHP were trained for 20,000 steps with 50 integration points and learning rates 0.01 and 0.1, respectively.

MIMIC-IV ICU

The MIMIC-IV search is implemented in scripts/tuning/tune_mimic_iv_icu_combined_local.sh on the top-116 ICU intervention vocabulary. Tuning runs use 20,000 optimization steps, 1,000-step warmup, and learning rate 0.001. Hawkes and standard neural runs use batch size 64 with 2-step gradient accumulation, and inhibitive runs use batch size 32 with 4-step accumulation.

Grids. The search sweeps Kb∈{5,8,10}K_b \in\{5, 8, 10\} for Spline Poisson, Kk∈{4,8,12}K_k \in\{4, 8, 12\} and Tf∈{2,5,10}T_f \in\{2, 5, 10\} for the constant-baseline spline-kernel Hawkes, and the Transformer backbone over L∈{2,4}L \in\{2, 4\}, H∈{2,4}H \in\{2, 4\}, d∈{32,64,128}d \in\{32, 64, 128\}, alongside the Poisson and exponential Hawkes baselines.

Reported configuration. The reported MIMIC-IV runs use Kb=20K_b = 20 and Tfactor=1.0T_{\mathrm{factor}} = 1.0 for Spline Poisson and Spline/Exp Hawkes, Tf=10T_f = 10 for the spline kernels, Tf=2T_f = 2 for NeuralSplineTPP, and a Transformer with L=4L = 4, H=4H = 4, d=64d = 64, dropout 0.1 and weight decay 10−410^{-4} for all Transformer models (IntensityFree: d=32d = 32, no weight decay). The kernel and head knot counts are selected as described in Appendix D.4.

Training Configuration

The results on the standard benchmarks and MIMIC-IV use the following configuration, with five seeds per model. All models are trained for 40,000 steps with early stopping on the validation log-likelihood, and the best validation checkpoint is restored before testing. The effective batch size is 128 throughout, using gradient accumulation where needed. Spline/Spline Hawkes and its inhibitive variant use 100 baseline knots, and the number of kernel/head spline knots of these models and of NeuralSplineTPP is selected on the validation log-likelihood from K∈{10,20,50,100,200,400}K \in\{10, 20, 50, 100, 200, 400\}. Numerically integrated neural models use per-interval integration with 10 points, and the inhibitive Hawkes processes use 100 grid points. Increasing the points of the inhibitive Hawkes processes to 10,000 (MIMIC-IV: 3,000) changes log⁡p(t)\log p(t) of seed 0 by at most 0.08 nats and leaves the type terms unchanged. Taobao and MIMIC-IV event times are stored in float64. The selected kernel knots are 20/10/400/400/20/20 for Spline/Spline Hawkes and 50/20/400/400/20/10 for NeuralSplineTPP on Taxi/StackOverflow/Retweet/Amazon/Taobao/MIMIC-IV. The inhibitive variant inherits them. Learning rates are 0.1 for Poisson and Hawkes models and 0.01 for neural models on the standard benchmarks, and 0.1 and 0.005 on MIMIC-IV. They were adjusted per model and dataset where training did not converge (e.g. inhibitive Hawkes 0.03–1.0, NHP 0.003 on MIMIC-IV, and NeuralSplineTPP 0.003 on Taobao). IntensityFree uses 0.01 throughout. On MIMIC-IV, NeuralConstTPP uses a Softplus head, as its exponential head diverged for individual seeds. All values are part of the released configurations (scripts/tuning/generate_gold_run_commands.sh). The UKB models use the configuration of Appendix § 11.§ 2. The knot-count sweep (Table 23, Figure 2) is run separately (20,000 steps on the standard benchmarks, learning rates 0.1 and 0.01 without per-model adjustment).

KKSpline/Spline Hawkes #paramsSpline/Spline Hawkes log p(t)p(t)Spline/Spline Hawkes log p(d∣t)p(d\mid t)Spline/Spline Hawkes Acc∣t\mid t (%)NeuralSplineTPP #paramsNeuralSplineTPP log p(t)p(t)NeuralSplineTPP log p(d∣t)p(d\mid t)NeuralSplineTPP Acc∣t\mid t (%)
10148,5913.69±.003.69\pm.00−2.60±.00-2.60\pm.0032.5±.032.5\pm.077,2203.99±.023.99\pm.02−2.49±.11-2.49\pm.1136.9±1.336.9\pm1.3
20285,4813.83±.003.83\pm.00−2.61±.00-2.61\pm.0032.9±.132.9\pm.1154,4403.99±.043.99\pm.04−2.61±.21-2.61\pm.2135.4±2.735.4\pm2.7
50696,1513.83±.003.83\pm.00−2.62±.00-2.62\pm.0032.8±.032.8\pm.0386,1003.97±.023.97\pm.02−2.82±.11-2.82\pm.1132.7±1.432.7\pm1.4
1001,380,6013.82±.003.82\pm.00−2.64±.00-2.64\pm.0032.8±.032.8\pm.0772,2003.97±.053.97\pm.05−2.89±.37-2.89\pm.3730.9±7.430.9\pm7.4
2002,749,5013.81±.003.81\pm.00−2.65±.00-2.65\pm.0032.7±.032.7\pm.01,544,4003.95±.083.95\pm.08−3.20±.50-3.20\pm.5023.5±10.023.5\pm10.0
4005,487,3013.79±.003.79\pm.00−2.68±.00-2.68\pm.0032.6±.032.6\pm.03,088,8003.98±.093.98\pm.09−3.03±.46-3.03\pm.4628.2±9.228.2\pm9.2

Table 23. Knot-count sweep on MIMIC-IV: Spline/Spline Hawkes (kernel knots KK; baseline fixed at 100 knots) vs. NeuralSplineTPP (head knots KK). Mean ±\pm std over 5 seeds. log⁡p(t)\log p(t) is flat in KK for both families, and log⁡p(d∣t)\log p(d\mid t) does not improve. #params: trainable parameters of the Hawkes model and of the NeuralSplineTPP head (its Transformer encoder adds 208,832).

Tuning Budget per Model Family

Table 6 summarizes the search effort per model family. The neural families received more search configurations than the Hawkes families (≈ 27 vs. ≈ 20 per standard benchmark dataset for the models in this paper), and at the validation-selected knot count the flexible Hawkes process and NeuralSplineTPP have about the same number of parameters on MIMIC-IV.

Search Philosophy

The tuning procedure is intentionally not a single flat Cartesian search over all models. Instead, the cheapest families are tuned first to identify reasonable optimization scales, transformer widths, and inter-event time scales; more expressive analytical or neural variants then inherit those choices and only spend their compute budget on model-specific parameters. This staged design substantially reduced the number of unstable runs on UKB and MIMIC-IV and made it feasible to compare analytical, numerical, and fully neural TPPs under a common experimental budget.

Compute Resources and Libraries

The experiments were run on a heterogeneous in-house GPU cluster with NVIDIA TITAN RTX (24 GB), NVIDIA L40S (48 GB), NVIDIA A100 (40 GB, both PCIe and SXM4), and NVIDIA H200 (141 GB) cards. MIMIC-IV and UKB runs defaulted to L40S or A100 to fit the larger transformer backbones, and part of the UKB experiments were run on the UK Biobank Research Analysis Platform. All runs used one GPU per job. No multi-node or multi-GPU training was used, and the longest single run took about 19 wall-clock hours.

Table 7 summarizes the number of training runs and the total wall-clock GPU time per dataset, separated into hyperparameter tuning and replica runs (5 seeds per model and dataset, including models and variants that are not reported in the tables). The tuning numbers are extracted from the WandB logs and include pilot configurations that were discarded. In total, these runs comprise approximately 1,210 training runs and ∼ 760 GPU-hours (∼ 32 GPU-days). We additionally reran the EasyTPP-internal baselines (THP, SAHP) using the original EasyTPP framework on the same TITAN RTX hardware for the equivalence check in Appendix § 8.§ 5; that retraining contributes another ∼80 runs (5 seeds × 2 models × 5 datasets, plus a small Taobao tuning sweep) and is not included in Table 7.

DatasetHyperparameter tuningReplica runs
RunsGPU-hRunsGPU-h
Standard benchmarks39614437073
MIMIC-IV ICU28475156
UK Biobank20528913398
Total629437578327

Table 7. Compute used for the experiments in this paper. Run counts and wall-clock GPU-hours are aggregated from WandB logs of the tuning sweeps and of the replica runs, which include the runs that produced the numbers in the tables. “Runs” counts only runs that reached the final evaluation step, and runtime is wall-clock time on a single GPU. Some tuning sweeps logged only metrics and not runtime. These are included in the run counts but not in the GPU-hour totals, which therefore slightly underestimate the true cost.

Failed and preempted runs are not accounted for, as the WandB logs we summarize only retain runs that reached the final evaluation step. The knot-count sweeps, the non-stationarity control (35 runs), and learning-rate adjustment runs used the same hardware and are not included in Table 7.

Our experiments used EasyTPP 0.2.1 for comparisons. It is available under the Apache-2.0 License. The exact software environment is available in our repository as a frozen uv.lock and pyproject.toml.

UK Biobank Data Preparation and Access

Access. Access to the UK Biobank requires an approved research application. We do not redistribute UK Biobank data, data derived from it, or models trained on it. The UKB results can therefore only be reproduced by approved UK Biobank users.

Preparation. We adapt the cohort and event construction of Shmatko et al. [2025] but restrict the vocabulary to first-occurrence ICD-10 disease tokens and a single terminal Death token, omitting the lifestyle and sex tokens of the original setup. Disease events are derived from ICD-10 codes truncated to three or four characters and joined to UK Biobank field IDs through the icd10_codes_mod.tsv mapping of the Delphi repository. After this mapping the vocabulary contains 1,256 disease codes plus a single Death token, giving D=1,257D = 1{,}257 event types after the data loader’s final remapping.

Cohort selection. A participant is included if they have a valid sex record. Records without a sex field are dropped, since sex is used both for cohort definition and as a sequence-start token in the original Delphi pipeline. We do not apply any further demographic or follow-up filters. Sex itself is not given to the models, and no other covariates are modeled. We list this as a limitation (§ 5).

Sequence construction and ordering. Each participant contributes exactly one sequence. Raw event timestamps are converted to days since the participant’s date of birth and sorted by (participant_id, time). Strictly simultaneous events are kept. TORCHPP handles them through causal masking.

Train/val/test split. The split is patient-level with fixed fractions 0.7/0.2/0.1. All events of a given participant are assigned to the same split, so there is no patient-level leakage across splits.

MIMIC-IV ICU Data Preparation

The MIMIC-IV ICU task is built from the MIMIC-IV-3.1 release [11, 12, 7], available to credentialed users via PhysioNet. The pipeline is split between scripts/data_prep/prepare_mimic_iv_datasets.py, which materializes the canonical sequences, and scripts/data_prep/analyze_mimic_iv_datasets.py, which computes vocabulary statistics. Both are orchestrated by scripts/data_prep/recreate_mimic_iv_benchmarks.py.

Source tables. We use icustays.csv.gz to define the per-stay observation windows [intime,outtime], and merge events from inputevents.csv.gz and procedureevents.csv.gz into a single event stream per ICU stay. Item metadata (label, abbreviation, category) is read from d_items.csv.gz, and no semantic re-grouping of items is applied.

Sequence unit. One sequence per ICU stay (stay_id), not per patient. A patient with multiple ICU stays therefore contributes multiple sequences. Within a stay, only events whose starttime falls inside [`intime, outtime\] are retained, and stays with fewer than two events are dropped. Event times are recorded to the minute. We add to each event an offset drawn uniformly within its recorded minute (truncated at outtime`, fixed seed) and sort the events by the resulting times. Event times are stored in float64.

Event vocabulary and OTHER class. The raw merged ICU vocabulary contains 486 itemids. We retain the top-116 itemids by training-set frequency, which together cover ≈95.1%\approx95.1\% of all training events, and remap the remaining 370 items to a single OTHER class. The task vocabulary therefore has D=117D = 117. The OTHER class is treated as a regular event type during training and evaluation; it inherits a synthetic metadata entry and is not masked out.

Sequence-length truncation. The data preparation step retains all events within the observation window; pre-truncation sequence lengths reach a 95th95^{\mathrm{th}} percentile of roughly 570 events. To bound the quadratic cost of the Transformer encoders and the all-pairs Hawkes kernels, sequences are right-truncated to 316 events (90th90^{\mathrm{th}} percentile of sequence lengths), reducing the average sequence length from 138.8 to 94.4 on the top-116 vocabulary.

Train/val/test split. The split is subject-level with fractions 0.8/0.1/0.1. All ICU stays for a given subject are placed in the same split. With the full vocabulary of 486 items, the splits contain 67,320/8,474/8,531 stays and 9,304,960/1,180,908/1,221,930 events; pre-truncation mean sequence lengths are 138.2, 139.4, and 143.2, respectively. The top-116 variant uses identical splits and stay assignments; only the vocabulary mapping differs.

Full Result Tables

Tables 8–14 report all metrics with standard deviations for the standard benchmarks, Table 15 for MIMIC-IV, and Table 16 for UKB. Log-likelihoods and Acc⁡∣t\operatorname{Acc}_{\mid t} are computed in closed form (or with the training-time integration for the numerically integrated models). In contrast, log⁡p(d)\log p(d), the marginal accuracy, and the RMSE require integration over the next-event time on a dense grid (§ 3.3). The RMSE is sensitive to this integration on datasets with heavy-tailed inter-event times (notably Retweet). On Taobao, whose test gaps are all below one, the RMSE does not separate the models.

ModelTaxiStackOverflowRetweetAmazonTaobao
Poisson processes
Homogeneous Poisson0.51±.00−0.86±.00−3.58±.00−0.36±.002.07±.00
Spline Poisson0.51±.00−0.85±.00−3.43±.00−0.35±.002.11±.00
Hawkes processes
Const./Exp Hawkes0.45±.00−0.76±.00−3.42±.00−0.10±.002.61±.00
Spline/Exp Hawkes0.45±.00−0.75±.00−3.38±.00−0.11±.002.61±.00
Const./Spline Hawkes0.50±.00−0.73±.00−3.40±.000.01±.002.65±.00
Spline/Spline Hawkes0.49±.00−0.73±.01−2.37±.001.73±.002.67±.00
Inhibitive Hawkes processes
Inhib. Const./Exp Hawkes0.54±.00−0.75±.00−3.41±.000.05±.022.62±.00
Inhib. Spline/Spline Hawkes0.68±.00−0.76±.00−1.29±.022.05±.012.66±.01
Neural TPPs
NeuralConstTPP0.55±.00−0.80±.01−3.42±.00−0.34±.002.07±.00
THP0.56±.00−0.80±.00−3.41±.02−0.31±.012.11±.00
SAHP0.64±.02−0.75±.00−3.38±.010.14±.062.66±.01
NHP0.60±.03−0.79±.01−3.51±.06−0.32±.032.56±.01
Flexible neural TPPs
IntensityFree0.73±.00−0.73±.01−3.66±.262.41±.012.73±.00
NeuralSplineTPP0.71±.00−0.72±.01−1.86±.012.56±.002.71±.00

Table 9. Time log-likelihood per event log⁡p(t∣H)\log p(t \mid\mathcal{H}) (nats, higher is better) on the standard benchmarks (Equation (12)). Mean ±\pm std over 5 seeds. Bold: best per dataset; underline: best classical (Poisson or Hawkes) model; ties at the displayed precision are all marked.

Table 10: Time-conditional type log-likelihood per event log p(d∣t,H)p(d \mid t, \mathcal{H}) at the true next-event time (nats, higher is better; Equation (13)). Mean ±\pm std over 5 seeds. Bold: best per dataset; underline: best classical (Poisson or Hawkes) model; ties at the displayed precision are all marked.

ModelTaxiStackOverflowRetweetAmazonTaobao
Poisson processes
Homogeneous Poisson−1.14±.00-1.14\pm.00−1.90±.00-1.90\pm.00−0.82±.00-0.82\pm.00−2.24±.00-2.24\pm.00−2.18±.00-2.18\pm.00
Spline Poisson−1.13±.00-1.13\pm.00−1.88±.00-1.88\pm.00−0.82±.00-0.82\pm.00−2.24±.00-2.24\pm.00−2.19±.00-2.19\pm.00
Hawkes processes
Const./Exp Hawkes−0.74±.00-0.74\pm.00−1.62±.00-1.62\pm.00−0.79±.00-0.79\pm.00−2.03±.03-2.03\pm.03−1.42±.00-1.42\pm.00
Spline/Exp Hawkes−0.74±.00-0.74\pm.00−1.60±.00-1.60\pm.00−0.78±.00-0.78\pm.00−2.04±.00-2.04\pm.00−1.43±.00-1.43\pm.00
Const./Spline Hawkes−0.71±.00-0.71\pm.00−1.60±.00-1.60\pm.00−0.79±.00-0.79\pm.00−1.96±.00-1.96\pm.00−1.38±.00-1.38\pm.00
Spline/Spline Hawkes−0.70±.00-0.70\pm.00−1.59±.00-1.59\pm.00−0.80±.00-0.80\pm.00−2.04±.00-2.04\pm.00−1.41±.00-1.41\pm.00
Inhibitive Hawkes processes
Inhib. Const./Exp Hawkes−0.33±.00-0.33\pm.00−1.60±.00-1.60\pm.00−0.78±.00-0.78\pm.00−2.06±.01-2.06\pm.01−1.43±.00-1.43\pm.00
Inhib. Spline/Spline Hawkes−0.28±.00-0.28\pm.00−1.60±.00-1.60\pm.00−0.80±.00-0.80\pm.00−2.17±.01-2.17\pm.01−1.47±.00-1.47\pm.00
Neural TPPs
NeuralConstTPP−0.29±.00-0.29\pm.00−1.59±.00-1.59\pm.00−0.78±.01-0.78\pm.01−1.99±.00-1.99\pm.00−1.43±.01-1.43\pm.01
THP−0.26±.00-0.26\pm.00−1.59±.00-1.59\pm.00−0.78±.00-0.78\pm.00−1.99±.00-1.99\pm.00−1.45±.01-1.45\pm.01
SAHP−0.24±.00-0.24\pm.00−1.57±.01-1.57\pm.01−0.79±.02-0.79\pm.02−2.20±.02-2.20\pm.02−1.49±.04-1.49\pm.04
NHP−0.24±.01-0.24\pm.01−1.60±.01-1.60\pm.01−0.81±.01-0.81\pm.01−2.06±.02-2.06\pm.02−1.51±.00-1.51\pm.00
Flexible neural TPPs
IntensityFree−0.29±.00-0.29\pm.00−1.60±.01-1.60\pm.01−0.79±.01-0.79\pm.01−2.04±.03-2.04\pm.03−1.42±.01-1.42\pm.01
NeuralSplineTPP−0.22±.00-0.22\pm.00−1.60±.01-1.60\pm.01−0.78±.00-0.78\pm.00−1.91±.01-1.91\pm.01−1.43±.01-1.43\pm.01

Table 10.

Table 12: Marginal top-1 type accuracy Acc (%, higher is better): arg maxdp(d∣H)\mathop{\rm arg\,max}_{d} p(d \mid\mathcal{H}) from the marginal type distribution (Equation (14)), computed by numerical integration. Grid-based metrics are evaluated on the first 1,000 test sequences. ∗^{*}: mean over 3–4 seeds, as the evaluation of a seed is missing or its RMSE diverged. Spline Hawkes on Amazon was not evaluated with the grid-based metrics. Mean ±\pm std over 5 seeds. Bold: best per dataset; underline: best classical (Poisson or Hawkes) model; ties at the displayed precision are all marked.

ModelTaxiStackOverflowRetweetAmazonTaobao
Poisson processes
Homogeneous Poisson44.29±\pm.0042.50±\pm.0053.92±\pm.0029.87±\pm.0043.57±\pm.00
Spline Poisson44.25±\pm.2842.50±\pm.0054.71±\pm.0129.87±\pm.0043.57±\pm.00
Hawkes processes
Const./Exp Hawkes85.29±\pm.4044.39±\pm.0657.62±\pm.1234.06±\pm1.2260.91±\pm.05
Spline/Exp Hawkes85.50±\pm.2844.45±\pm.0358.71±\pm.1133.21±\pm.0260.79±\pm.06
Const./Spline Hawkes80.32±\pm.1344.29±\pm.0557.44±\pm.1335.17±\pm.0661.20±\pm.08
Spline/Spline Hawkes79.27±\pm.0744.34±\pm.0257.71±\pm.13–61.19±\pm.04
Inhibitive Hawkes processes
Inhib. Const./Exp Hawkes90.55±\pm.02<u>44.81±\pm.03</u>57.88±\pm.2828.57±\pm.4460.92±\pm.04
Inhib. Spline/Spline Hawkes<u>91.25±\pm.00</u><u>44.48±\pm.03</u>58.45±\pm.1335.06±\pm.3460.60±\pm.13
Neural TPPs
NeuralConstTPP91.44±\pm.0645.25±\pm.1058.58±\pm1.2435.62±\pm.0560.23±\pm.16
THP91.53±\pm.0545.17±\pm.1759.33±\pm.3135.67±\pm.0960.33±\pm.17
SAHP91.47±\pm.0944.59±\pm.0558.31±\pm2.4335.47±\pm.2059.90±\pm.39
NHP91.53±\pm.1444.58±\pm.0256.30±\pm2.3933.31±\pm1.1160.63±\pm.12
Flexible neural TPPs
IntensityFree91.49±\pm.0745.28±\pm.13 ∗^{*}58.69±\pm.9434.87±\pm.6660.14±\pm.27
NeuralSplineTPP91.52±\pm.0743.84±\pm.0559.01±\pm.6035.87±\pm.17 ∗^{*}60.09±\pm.11

Table 12.

Table 13: Marginal type log-likelihood per event log⁡p(d∣H)\log p(d \mid\mathcal{H}) (nats, higher is better; Equation (14)), computed by numerical integration. Grid-based metrics are evaluated on the first 1,000 test sequences. ∗^{*}: mean over 3–4 seeds, as the evaluation of a seed is missing or its RMSE diverged. Spline/Spline Hawkes on Amazon was not evaluated with the grid-based metrics. Mean ±\pm std over 5 seeds. Bold: best per dataset; underline: best classical (Poisson or Hawkes) model; ties at the displayed precision are all marked.

ModelTaxiStackOverflowRetweetAmazonTaobao
Poisson processes
Homogeneous Poisson−1.14±\pm.00−1.90±\pm.00−0.82±\pm.00−2.24±\pm.00−2.18±\pm.00
Spline Poisson−1.14±\pm.00−1.88±\pm.00−0.82±\pm.00−2.24±\pm.00−2.19±\pm.00
Hawkes processes
Const./Exp Hawkes−0.81±\pm.00−1.64±\pm.00−0.79±\pm.00−2.03±\pm.02−1.42±\pm.00
Spline/Exp Hawkes−0.80±\pm.00−1.63±\pm.00−0.78±\pm.00−2.03±\pm.00−1.42±\pm.00
Const./Spline Hawkes−0.83±\pm.00−1.70±\pm.00−0.79±\pm.00<u>−2.02±\pm.00</u>−1.39±\pm.00
Spline/Spline Hawkes−0.83±\pm.00−1.66±\pm.00−0.79±\pm.00–<u>−1.39±\pm.00</u>
Inhibitive Hawkes processes
Inhib. Const./Exp Hawkes−0.40±\pm.00<u>−1.61±\pm.00</u>−0.78±\pm.00−2.12±\pm.02−1.43±\pm.00
Inhib. Spline/Spline Hawkes<u>−0.32±\pm.00</u>−1.63±\pm.00−0.79±\pm.00−2.05±\pm.02−1.40±\pm.00
Neural TPPs
NeuralConstTPP−0.29±\pm.00−1.59±\pm.00−0.78±\pm.01−1.99±\pm.00−1.43±\pm.01
THP−0.29±\pm.00−1.59±\pm.00−0.78±\pm.00−1.99±\pm.00−1.44±\pm.00
SAHP−0.29±\pm.00−1.61±\pm.01−0.79±\pm.02−2.02±\pm.01−1.47±\pm.02
NHP−0.30±\pm.01−1.63±\pm.01−0.81±\pm.01−2.07±\pm.02−1.46±\pm.00
Flexible neural TPPs
IntensityFree−0.29±\pm.00−1.60±\pm.00 ∗^{*}−0.79±\pm.01−2.04±\pm.03−1.42±\pm.01
NeuralSplineTPP−0.30±\pm.00−1.70±\pm.01−0.78±\pm.00−1.99±\pm.01 ∗^{*}−1.43±\pm.00

Table 13.

Table 14: Next-event time RMSE (lower is better) using the predictive mean E[t∣H]\mathbb{E}[t \mid\mathcal{H}], in each dataset’s native EasyTPP time unit (not comparable across datasets). {sup}‡IntensityFree: the closed-form mean of its log-normal mixture is dominated by components with large variance and overflows for individual seeds, so we report the median predictor instead and exclude it from the bold marking. Grid-based metrics are evaluated on the first 1,000 test sequences. {sup}∗`: mean over 3–4 seeds, as the evaluation of a seed is missing or its RMSE diverged. Spline/Spline Hawkes on Amazon was not evaluated with the grid-based metrics. Mean ± std over 5 seeds. Bold: best per dataset; underline: best classical (Poisson or Hawkes) model; ties at the displayed precision are all marked.

ModelTaxiStackOverflowRetweetAmazonTaobao
Poisson processes
Homogeneous Poisson0.30±.000.30\pm.001.07±.001.07\pm.0017.82±.0117.82\pm.010.34±.000.34\pm.000.13±.000.13\pm.00
Spline Poisson0.30±.000.30\pm.001.07±.001.07\pm.0016.22±.0116.22\pm.010.34±.000.34\pm.000.13±.000.13\pm.00
Hawkes processes
Const./Exp Hawkes0.31±.000.31\pm.001.03±.001.03\pm.0016.65±.0116.65\pm.010.36±.010.36\pm.010.13±.000.13\pm.00
Spline/Exp Hawkes0.31±.000.31\pm.001.03±.001.03\pm.0015.83±.0115.83\pm.010.34±.000.34\pm.000.13±.000.13\pm.00
Const./Spline Hawkes0.32±.000.32\pm.001.04±.001.04\pm.0016.63±.0116.63\pm.010.37±.000.37\pm.000.13±.000.13\pm.00
Spline/Spline Hawkes0.35±.000.35\pm.001.03±.001.03\pm.0020.67±.2620.67\pm.26–0.13±.000.13\pm.00
Inhibitive Hawkes processes
Inhib. Const./Exp Hawkes0.30±.000.30\pm.001.02±.001.02\pm.0016.54±.0616.54\pm.060.36±.010.36\pm.010.13±.000.13\pm.00
Inhib. Spline/Spline Hawkes0.29±.000.29\pm.001.03±.001.03\pm.0016.35±.0216.35\pm.020.35±.000.35\pm.000.13±.000.13\pm.00
Neural TPPs
NeuralConstTPP0.28±.000.28\pm.001.03±.011.03\pm.0116.15±.0416.15\pm.040.33±.000.33\pm.000.13±.000.13\pm.00
THP0.28±.000.28\pm.001.03±.001.03\pm.0016.04±.1616.04\pm.160.33±.000.33\pm.000.13±.000.13\pm.00
SAHP0.28±.000.28\pm.001.04±.001.04\pm.0016.11±.1316.11\pm.130.34±.010.34\pm.010.13±.000.13\pm.00
NHP0.28±.000.28\pm.001.04±.011.04\pm.0117.05±.6517.05\pm.650.34±.000.34\pm.000.13±.000.13\pm.00
Flexible neural TPPs
IntensityFree0.29±.000.29\pm.001.09±.021.09\pm.02{sup}∗‡`16.72±.3316.72\pm.33{sup}‡`0.40±.000.40\pm.00{sup}‡`0.13±.000.13\pm.00{sup}‡`
NeuralSplineTPP0.28±.000.28\pm.001.05±.011.05\pm.0116.42±.1716.42\pm.170.33±.000.33\pm.00{sup}∗`0.13±.000.13\pm.00

Table 14.

Discretized Time Likelihood

For an event at time tit_i, a previous event at ti−1t_{i-1}, and a bin width Δ\Delta, let b=(⌊ti/Δ⌋+1)Δb = (\lfloor t_i/\Delta\rfloor+ 1)\Delta be the end of the bin that contains tit_i and a=max⁡(⌊ti/Δ⌋Δ,ti−1)a = \max(\lfloor t_i/\Delta\rfloor\Delta, t_{i-1}) its start, truncated at the previous event. The binned time log-likelihood is

log⁡Pr⁡(ti∈[a,b)∣H)=log⁡(S∗(a)−S∗(b))=−Λ∗(ti−1,a)+log⁡(1−e−(Λ∗(ti−1,b)−Λ∗(ti−1,a))),\log\Pr(t_i \in[a,b) \mid\mathcal{H}) = \log(S^*(a)-S^*(b)) = -\Lambda^*(t_{i-1},a)+\log\left(1-e^{-(\Lambda^*(t_{i-1},b)-\Lambda^*(t_{i-1},a))}\right),

with the survival function S∗S^* of the next event after ti−1t_{i-1} (Appendix B). We compute it from the same checkpoints, test sequences, and (closed-form or numerical) cumulative intensities as the continuous log p(t)p(t). Unlike the continuous density, it is bounded by zero, so a model cannot gain by placing narrow peaks on the grid points. Its level depends on Δ\Delta, and only differences between models or knot counts at the same Δ\Delta are meaningful. Bins start at multiples of Δ\Delta on the time axis of each dataset. On Taobao, whose gaps rather than time stamps lie on a grid, bins start at the previous event. On MIMIC-IV, the 1-minute bin of an event is its recorded minute. As a check, we also use bins centered on the grid points.

Table 18 lists the recorded resolution of each dataset and the bins we use. StackOverflow is distributed in single precision. Table 19 shows that the gain from more knots survives binning only on Amazon. On Retweet, a continuous gain of up to 1.5 nats turns into a small loss at the recorded resolution. On MIMIC-IV, both are flat. Centered bins do not change this conclusion. Table 20 reports the binned likelihood for the models of the main tables. On Taxi, StackOverflow, and Amazon, the order of the best models is essentially unchanged. On Retweet, the continuous lead of the high-knot spline models disappears. On Taobao, the binned and the continuous likelihood rank the models the same. On MIMIC-IV, the flexible neural models keep their lead, which comes from the events recorded in the same minute as the previous event (Appendix J).

DatasetRecorded resolutionBinsTied (%)
Taxi1 s1 s, 10 s0.0
StackOverflowfloat32 (2−132^{-13})2−132^{-13}, 10⋅2−1310 \cdot2^{-13}0.5
Retweetinteger time stamps1, 107.4
Amazonnone (minimum gap 0.01)0.01, 0.10.0
Taobaogaps of 10−410^{-4}10−410^{-4}, 10−310^{-3}0.0
MIMIC-IV1 min (dequantized)1 min, 5 min0.0

Table 18. Time resolution of the evaluated event times (first 1,000 test sequences, MIMIC-IV truncated to 316 events) and the bin widths used for the binned time likelihood, in the time unit of each dataset. Tied: fraction of evaluated events with a zero gap to the previous event.

DatasetModellog⁡p(t)\log p(t), continuousΔ\Delta binned
K=20K = 20Δ\Delta to K=400K = 400recordedcoarsercentered
TaxiNeuralSplineTPP0.73-0.05-0.05-0.04-0.05
Spline/Spline Hawkes0.49+0.01+0.01+0.01+0.01
StackOverflowNeuralSplineTPP-0.72-0.04-0.05-0.06-0.09
Spline/Spline Hawkes-0.73-0.03-0.04-0.04-0.04
RetweetNeuralSplineTPP-3.32+1.47-0.06+0.00-0.28
Spline/Spline Hawkes-3.36+0.99-0.13-0.11-0.12
AmazonNeuralSplineTPP1.41+1.15+0.93+0.56–
Spline/Spline Hawkes0.43+1.30+1.12+0.83–
TaobaoNeuralSplineTPP2.70+0.01-0.01+0.02+0.01
Spline/Spline Hawkes2.67-0.04-0.04-0.03-0.04
MIMIC-IVNeuralSplineTPP3.99-0.01-0.02-0.04-0.03
Spline/Spline Hawkes3.83-0.05-0.04-0.04-0.03

Table 19. Change of log⁡p(t)\log p(t) from K=20K = 20 to K=400K = 400 knots (nats, mean over 5 seeds), continuous and binned at the recorded resolution, at a ten-fold (MIMIC-IV: five-fold) coarser bin, and with bins centered on the grid points instead of starting at them (Table 18). Same checkpoints as Table 22.

ModelTaxiStackOverflowRetweetAmazonTaobaoMIMIC-IV
Poisson processes
Homogeneous Poisson−7.68±.00−9.86±.00−3.62±.00−4.96±.00−7.14±.00−4.90±.01
Spline Poisson−7.68±.00−9.85±.00−3.48±.00−4.96±.00−7.10±.00−4.87±.00
Hawkes processes
Const./Exp Hawkes−7.74±.00−9.76±.00−3.47±.00−4.70±.00−6.60±.00−4.13±.00
Spline/Exp Hawkes−7.74±.00−9.75±.00−3.43±.00−4.71±.00−6.60±.00−4.12±.00
Const./Spline Hawkes−7.69±.00<u>−9.72</u>±.00−3.46±.00−4.59±.00−6.56±.00−3.73±.00
Spline/Spline Hawkes−7.70±.00<u>−9.72</u>±.01−3.55±.00−3.09±.00<u>−6.54</u>±.00−3.67±.00
Inhibitive Hawkes processes
Inhib. Const./Exp Hawkes−7.65±.00−9.75±.00−3.46±.00−4.55±.02−6.60±.00<u>−3.64</u>±.00
Inhib. Spline/Spline Hawkes<u>−7.51</u>±.00−9.76±.00−3.56±.00<u>−2.80</u>±.01−6.56±.01−3.68±.04
Neural TPPs
NeuralConstTPP−7.63±.00−9.80±.01−3.47±.00−4.95±.00−7.14±.00−4.42±.02
THP−7.63±.00−9.80±.00−3.46±.01−4.92±.01−7.10±.00−4.24±.01
SAHP−7.55±.02−9.74±.00−3.43±.01−4.46±.06−6.55±.01−3.88±.01
NHP−7.59±.03−9.79±.01−3.54±.06−4.93±.03−6.65±.01−3.92±.01
Flexible neural TPPs
IntensityFree−7.46±.00−9.70±.00−3.51±.02−2.43±.01−6.49±.01−3.43±.02
NeuralSplineTPP−7.48±.00−9.71±.01−3.52±.01−2.32±.00−6.50±.00−3.39±.02

Table 20. Binned time log-likelihood log⁡Pr⁡(t∈bin)\log\Pr(t \in\text{bin}) at the recorded resolution (nats, higher is better; bins in Table 18) for the models of the main tables. Only differences within a column are meaningful, as the level depends on the bin width. On Retweet the history-aware models are within a few tenths of a nat, while the continuous log⁡p(t)\log p(t) (Table 9) spreads over more than two nats. Mean ± std over 5 seeds. Bold: best per dataset; <u>underline</u>: best classical (Poisson or Hawkes) model; ties at the displayed precision are all marked.

Controls on MIMIC-IV

Simultaneous events. Of the evaluated events, 41.6% are recorded in the same minute as the previous event (Table 21). On these, NeuralSplineTPP and THP are far more accurate than the Hawkes processes (NeuralSplineTPP: 59.5%, Spline/Spline Hawkes: 47.4%). On the other events, NeuralSplineTPP and Spline/Spline Hawkes differ by 0.6 percentage points in Acc∣t|t and 0.02 nats in log⁡p(d∣t)\log p(d|t), and Spline/Spline Hawkes has a 0.09 nats higher log⁡p(t)\log p(t). In binned time at one minute (Table 21), NeuralSplineTPP leads Spline/Spline Hawkes by 0.28 nats. It leads by 0.79 nats on the same-minute events and trails by 0.09 nats on the other events, where Inhib. Spline/Spline Hawkes has the best binned log⁡p(t)\log p(t).

Modellog p(t)p(t) alllog p(t)p(t) otherBinned log p(t)p(t) allBinned log p(t)p(t) otherBinned log p(t)p(t) samelog p(d∣t)p(d|t) alllog p(d∣t)p(d|t) otherlog p(d∣t)p(d|t) sameAcc∣t\mid t (%) allAcc∣t\mid t (%) otherAcc∣t\mid t (%) same
Homogeneous Poisson2.60±.012.60\pm.011.71±.041.71\pm.04−4.90±.01-4.90\pm.01−5.57±.04-5.57\pm.04−3.97±.04-3.97\pm.04−3.78±.02-3.78\pm.02−3.91±.02-3.91\pm.02−3.60±.02-3.60\pm.0213.6±.013.6\pm.012.7±.012.7\pm.014.7±.014.7\pm.0
Spline Poisson2.64±.002.64\pm.001.79±.031.79\pm.03−4.87±.00-4.87\pm.00−5.49±.03-5.49\pm.03−3.99±.03-3.99\pm.03−3.71±.01-3.71\pm.01−3.84±.01-3.84\pm.01−3.53±.01-3.53\pm.0113.6±.013.6\pm.012.7±.012.7\pm.014.7±.014.7\pm.0
Const./Exp Hawkes3.38±.003.38\pm.002.05±.002.05\pm.00−4.13±.00-4.13\pm.00−5.24±.00-5.24\pm.00−2.57±.00-2.57\pm.00−2.71±.00-2.71\pm.00−3.12±.00-3.12\pm.00−2.13±.00-2.13\pm.0030.9±.130.9\pm.120.1±.120.1\pm.146.2±.246.2\pm.2
Spline/Exp Hawkes3.39±.003.39\pm.002.05±.002.05\pm.00−4.12±.00-4.12\pm.00−5.23±.00-5.23\pm.00−2.57±.00-2.57\pm.00−2.70±.00-2.70\pm.00−3.12±.00-3.12\pm.00−2.12±.00-2.12\pm.0030.9±.030.9\pm.020.1±.120.1\pm.146.2±.146.2\pm.1
Const./Spline Hawkes3.78±.003.78\pm.002.11±.002.11\pm.00−3.73±.00-3.73\pm.00−5.17±.00-5.17\pm.00−1.71±.01-1.71\pm.01−2.62±.00-2.62\pm.00−2.95±.00-2.95\pm.00−2.15±.00-2.15\pm.0032.7±.032.7\pm.022.4±.022.4\pm.047.3±.047.3\pm.0
Spline/Spline Hawkes3.83±.003.83\pm.002.15±.002.15\pm.00−3.67±.00-3.67\pm.00−5.13±.00-5.13\pm.00−1.61±.00-1.61\pm.00−2.61±.00-2.61\pm.00−2.93±.00-2.93\pm.00−2.17±.00-2.17\pm.0032.9±.132.9\pm.122.5±.122.5\pm.147.4±.147.4\pm.1
Inhib. Const./Exp Hawkes3.87±.003.87\pm.002.10±.002.10\pm.00−3.64±.00-3.64\pm.00−5.19±.00-5.19\pm.00−1.47±.01-1.47\pm.01−2.67±.01-2.67\pm.01−3.07±.01-3.07\pm.01−2.12±.00-2.12\pm.0032.1±.232.1\pm.221.3±.321.3\pm.347.3±.147.3\pm.1
Inhib. Spline/Spline Hawkes3.83±.043.83\pm.042.24±.012.24\pm.01−3.68±.04-3.68\pm.04−5.04±.02-5.04\pm.02−1.77±.07-1.77\pm.07−2.68±.09-2.68\pm.09−3.04±.12-3.04\pm.12−2.19±.03-2.19\pm.0333.5±.333.5\pm.322.9±.522.9\pm.548.3±.148.3\pm.1
NeuralConstTPP3.08±.023.08\pm.022.18±.012.18\pm.01−4.42±.02-4.42\pm.02−5.10±.01-5.10\pm.01−3.47±.05-3.47\pm.05−2.65±.01-2.65\pm.01−3.07±.01-3.07\pm.01−2.06±.01-2.06\pm.0132.6±.332.6\pm.319.5±.219.5\pm.251.1±.651.1\pm.6
THP3.26±.013.26\pm.012.11±.012.11\pm.01−4.24±.01-4.24\pm.01−5.16±.01-5.16\pm.01−2.95±.03-2.95\pm.03−2.36±.01-2.36\pm.01−2.87±.00-2.87\pm.00−1.65±.01-1.65\pm.0137.6±.137.6\pm.121.9±.121.9\pm.159.6±.259.6\pm.2
SAHP3.66±.013.66\pm.012.00±.012.00\pm.01−3.88±.01-3.88\pm.01−5.28±.01-5.28\pm.01−1.91±.02-1.91\pm.02−2.74±.03-2.74\pm.03−3.27±.03-3.27\pm.03−2.00±.05-2.00\pm.0535.3±.435.3\pm.421.0±.421.0\pm.455.6±.955.6\pm.9
NHP3.60±.013.60\pm.011.86±.011.86\pm.01−3.92±.01-3.92\pm.01−5.42±.01-5.42\pm.01−1.80±.02-1.80\pm.02−2.98±.01-2.98\pm.01−2.97±.01-2.97\pm.01−2.99±.02-2.99\pm.0221.9±.221.9\pm.222.0±.222.0\pm.221.6±.721.6\pm.7
IntensityFree4.06±.024.06\pm.022.05±.022.05\pm.02−3.43±.02-3.43\pm.02−5.23±.02-5.23\pm.02−0.90±.02-0.90\pm.02−2.61±.01-2.61\pm.01−3.09±.01-3.09\pm.01−1.93±.02-1.93\pm.0234.3±.334.3\pm.319.1±.319.1\pm.355.7±.455.7\pm.4
NeuralSplineTPP4.05±.024.05\pm.022.06±.022.06\pm.02−3.39±.02-3.39\pm.02−5.22±.02-5.22\pm.02−0.83±.02-0.83\pm.02−2.39±.01-2.39\pm.01−2.91±.01-2.91\pm.01−1.65±.01-1.65\pm.0138.3±.138.3\pm.123.1±.123.1\pm.159.5±.259.5\pm.2

Table 21. MIMIC-IV results split by whether the evaluated event was recorded in the same minute as the previous event (same, 41.6% of the evaluated events) or not (other). Same checkpoints and test sequences as Table 15. Binned: log⁡Pr⁡(t∈bin)\log\Pr(t \in\text{bin}) at one minute (Appendix I). Mean ±\pm std over 5 seeds. Bold: best per column; underline: best classical (Poisson or Hawkes) model; ties at the displayed precision are all marked.

Flexibility and capacity. Table 23 sweeps the number of spline knots of Spline/Spline Hawkes and NeuralSplineTPP on MIMIC-IV from 10 to 400. For both families, log⁡p(t)\log p(t) is flat for K≥20K \ge20. The log⁡p(d∣t)\log p(d|t) of Spline/Spline Hawkes decreases slightly with KK, and that of NeuralSplineTPP decreases strongly, with a large spread over seeds. The type-conditional advantage of NeuralSplineTPP is therefore largest at K=10K = 10. We attribute the decrease of log⁡p(d∣t)\log p(d|t) with growing KK to the higher variance of the ratio λd∗(t)/∑d′λd′∗(t)\lambda_d^*(t)/\sum_{d'} \lambda_{d'}^*(t) as the intensities become spikier.

Non-stationarity. An alternative explanation for the MIMIC-IV gap is a drift of the dynamics over the ICU stay, which a stationary kernel cannot represent. We therefore built non-stationary (NS) variants of Spline/Exp and Spline/Spline Hawkes whose kernel parameters are themselves linear splines over absolute time within the stay (Kns∈{3,5}K_{\mathrm{ns}} \in\{3,5\} knots). The cumulative intensity stays in closed form, and the model reduces exactly to the stationary kernel for constant heights. We trained 7 settings with 5 seeds each (Table 24). No NS setting improves the joint log-likelihood over the stationary model with the nearest knot count by more than 0.002 nats. The best NS setting, selected on validation likelihood, lowers it by 0.12 nats and widens the gap to NeuralSplineTPP from 0.45 to 0.57 nats. The largest improvement of any component is 0.03 nats in log⁡p(t)\log p(t), for KΔt=12K_{\Delta t} = 12 against K=10K = 10. Within-stay non-stationarity therefore does not explain the gap.

ModelKΔtK_{\Delta t}KnsK_{ns}log⁡p(t)\log p(t)log⁡p(d∣t)\log p(d\mid t)Acc∣t\mid t (%)
Spline/Exp Hawkes–13.39±.003.39 \pm.00−2.70±.00-2.70 \pm.0030.9±.030.9 \pm.0
+ NS kernel–53.36±.003.36 \pm.00−2.69±.00-2.69 \pm.0031.1±.031.1 \pm.0
Spline/Spline Hawkes + NS kernel1233.72±.003.72 \pm.00−2.63±.00-2.63 \pm.0032.9±.032.9 \pm.0
Spline/Spline Hawkes + NS kernel1253.71±.003.71 \pm.00−2.63±.00-2.63 \pm.0033.0±.033.0 \pm.0
Spline/Spline Hawkes2013.83±.003.83 \pm.00−2.61±.00-2.61 \pm.0032.9±.132.9 \pm.1
+ NS kernel2033.76±.003.76 \pm.00−2.66±.00-2.66 \pm.0032.9±.132.9 \pm.1
+ NS kernel2053.74±.003.74 \pm.00−2.66±.00-2.66 \pm.0032.9±.132.9 \pm.1
Spline/Spline Hawkes10013.82±.003.82 \pm.00−2.64±.00-2.64 \pm.0032.8±.032.8 \pm.0
+ NS kernel10033.73±.003.73 \pm.00−2.67±.00-2.67 \pm.0032.9±.032.9 \pm.0
+ NS kernel10053.71±.003.71 \pm.00−2.68±.00-2.68 \pm.0032.9±.032.9 \pm.0

Table 24. Non-stationary Hawkes control on MIMIC-IV. The kernel heights of the NS variants are themselves linear splines over absolute stay time with KnsK_{ns} knots (Kns=1K_{ns} = 1: the stationary model); KΔtK_{\Delta t}: knots of the kernel over elapsed time. Mean ±\pm std over 5 seeds. The NS runs use the configuration of Appendix D.4 with learning rate 0.05 for the NS Spline/Spline kernels. The stationary Spline/Spline rows are taken from the knot sweep (Table 23). At KΔt=12K_{\Delta t} = 12, their nearest counterparts are K=10K = 10 and 20 there. No non-stationary cell improves the joint log-likelihood of its nearest stationary counterpart by more than 0.002 nats, and all remain below NeuralSplineTPP (log⁡p(t)=4.05\log p(t) = 4.05, log⁡p(d∣t)=−2.39\log p(d\mid t) = -2.39).

Calibration

As a goodness-of-fit check beyond likelihood and accuracy, Table 25 reports time-rescaling calibration [4] and the calibration of the type distribution for four representative models with a closed-form cumulative intensity (so that the compensator is exact). The flexible models are best calibrated in time on four of the five standard datasets. On MIMIC-IV, NeuralSplineTPP is well calibrated in time, whereas the other models are not.

References

  1. [1]Emmanuel Bacry, Martin Bompaire, Philip Deegan, Stéphane Gäffas, and Søren V. Poulsen. Tick: a Python library for statistical learning, with an emphasis on Hawkes processes and time-dependent models. Journal of Machine Learning Research, 18(214):1–5, 2018.
  2. [2]Tanguy Bosser and Souhaib Ben Taieb. On the Predictive Accuracy of Neural Temporal Point Process Models for Continuous-time Event Data. Transactions on Machine Learning Research, 2023.arxiv.org/abs/2306.17066
  3. [3]Pierre Brémaud and Laurent Massoullié. Stability of nonlinear Hawkes processes. The Annals of Probability, 24(3):1563–1588, 1996.DOI
  4. [4]Emery N. Brown, Riccardo Barbieri, Valentina Ventura, Robert E. Kass, and Loren M. Frank. The time-rescaling theorem and its application to neural spike train data analysis. Neural Computation, 14(2):325–346, 2002.pubmed.ncbi.nlm.nih.gov/11802915
  5. [5]Yuxin Chang, Alex Boyd, Cao Xiao, Taha Kass-Hout, Parminder Bhatia, Padhraic Smyth, and Andrew Warrington. Deep Continuous-Time State-Space Models for Marked Event Sequences. Advances in Neural Information Processing Systems, 2025.pubmed.ncbi.nlm.nih.gov/42158290
  6. [6]Nan Du, Hanjun Dai, Rakshit Trivedi, Utkarsh Upadhyay, Manuel Gomez-Rodriguez, and Le Song. Recurrent Marked Temporal Point Processes: Embedding Event History to Vector. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 1555–1564, 2016.DOI
  7. [7]Ary L. Goldberger, Luis A. N. Amaral, Leon Glass, Jeffrey M. Hausdorff, Plamen Ch. Ivanov, Roger G. Mark, Joseph E. Mietus, George B. Moody, Chung-Kang Peng, and H. Eugene Stanley. PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals. Circulation, 101(23):e215–e220, 2000.
  8. [8]A. G. Hawkes. Spectra of some self-exciting and mutually exciting point processes. Biometrika, 58(1):83–90, 1971.DOI
  9. [9]Valerie Isham and Mark Westcott. A self-correcting point process. Stochastic Processes and their Applications, 8:335–347, 1979.DOI
  10. [11]Alistair E. W. Johnson, Lucas Bulgarelli, Lu Shen, Alvin Gayles, Ayad Shammout, Steven Horng, Tom J. Pollard, Sicheng Hao, Benjamin Moody, Brian Gow, Li-wei H. Lehman, Leo A. Celi, and Roger G. Mark. MIMIC-IV, a freely accessible electronic health record dataset. Scientific Data, 10(1):1, 2023.DOI
  11. [12]Alistair Johnson, Lucas Bulgarelli, Tom Pollard, Brian Gow, Benjamin Moody, Steven Horng, Leo Anthony Celi, and Roger Mark. MIMIC-IV (version 3.1). PhysioNet, 2024. https://doi.org/10.13026/kpb9-mt58.
  12. [13]Gavin Kerrigan, Kai Nelson, and Padhraic Smyth. EventFlow: Forecasting Temporal Point Processes with Flow Matching. International Conference on Artificial Intelligence and Statistics (AISTATS), 2026.
  13. [14]Jure Leskovec and Andrej Krevl. SNAP Datasets: Stanford Large Network Dataset Collection. 2014. http://snap.stanford.edu/data.
  14. [15]Yikuan Li, Shishir Rao, Jose Roberto Ayala Solares, Abdelaali Hassaine, Rema Ramakrishnan, Dexter Canoy, Yajie Zhu, Kazem Rahimi, and Gholamreza Salimi-Khorshidi. BEHRT: Transformer for Electronic Health Records. Scientific Reports, 10:7155, 2020.
  15. [16]Angelica Liguori, Luciano Caroprese, Marco Minici, Bruno Veloso, Francesco Spinnato, Mirco Nanni, Giuseppe Manco, and João Gama. Modeling events and interactions through temporal processes: A survey. Neurocomputing, 653:131191, 2025.arxiv.org/abs/2303.06067
  16. [17]Haitao Lin, Lirong Wu, Guojiang Zhao, Pai Liu, and Stan Z. Li. Exploring generative neural temporal point process. Transactions on Machine Learning Research, 2022.arxiv.org/abs/2208.01874
  17. [18]Haitao Lin, Cheng Tan, Lirong Wu, Zicheng Liu, Zhangyang Gao, and Stan Z. Li. An Extensive Survey With Empirical Studies on Deep Temporal Point Process. IEEE Transactions on Knowledge and Data Engineering, 37(4):1599–1619, 2025.DOI
  18. [19]David Lüdke, Marin Biloš, Oleksandr Schur, Marten Lienen, and Stephan Günnemann. Add and Thin: Diffusion for Temporal Point Processes. Advances in Neural Information Processing Systems, 2023.arxiv.org/abs/2311.01139
  19. [20]David Lüdke, Marten Lienen, Marcel Kollovieh, and Stephan Günnemann. Edit-Based Flow Matching for Temporal Point Processes. International Conference on Learning Representations, 2026.arxiv.org/abs/2510.06050
  20. [21]Hongyuan Mei and Jason Eisner. The Neural Hawkes Process: A Neurally Self-Modulating Multivariate Point Process. Advances in Neural Information Processing Systems, 2017.arxiv.org/abs/1612.09328
  21. [22]Amartya Mukherjee, Ruizhi Deng, He Zhao, Yuzhen Mao, Leonid Sigal, and Frederick Tung. Adiff4TTP: Asynchronous Diffusion Models for Temporal Point Processes. Transactions on Machine Learning Research, 2026.arxiv.org/abs/2504.20411
  22. [23]Jianmo Ni, Jiacheng Li, and Julian McAuley. Justifying Recommendations using Distantly-Labeled Reviews and Fine-Grained Aspects. In Proceedings of EMNLP-IJCNLP, pages 188–197, 2019.DOI
  23. [24]Yosihiko Ogata. On Lewis’ simulation method for point processes. IEEE Transactions on Information Theory, 27(1):23–31, 1981.DOI
  24. [25]Yosihiko Ogata. Statistical models for earthquake occurrences and residual analysis for point processes. Journal of the American Statistical Association, 83(401):9–27, 1988.DOI
  25. [27]Laila Rasmy, Yang Xiang, Ziqian Xie, Cui Tao, and Degui Zhi. Med-BERT: pretrained contextualized embeddings on large-scale structured electronic health records for disease prediction. npj Digital Medicine, 4(1):86, 2021.DOI
  26. [28]Paweł Renc, Yugang Jia, Anthony E. Samir, Jarosław Wąs, Quanzheng Li, David W. Bates, and Arkadiusz Sitek. Zero shot health trajectory prediction using transformer. npj Digital Medicine, 7:256, 2024.DOI
  27. [29]Oleksandr Schur, Marin Biloš, and Stephan Günnemann. Intensity-Free Learning of Temporal Point Processes. International Conference on Learning Representations, 2020.
  28. [30]Oleksandr Schur, Nicholas Gao, Marin Biloš, and Stephan Günnemann. Fast and Flexible Temporal Point Processes with Triangular Maps. Advances in Neural Information Processing Systems, 2020.
  29. [31]Oleksandr Schur, Ali Caner Türkmen, Tim Januschowski, and Stephan Günnemann. Neural Temporal Point Processes: A Review. In Proceedings of the 30th International Joint Conference on Artificial Intelligence (IJCAI), pages 4585–4593, 2021.
  30. [32]Artem Shmatko, Alexander Wolfgang Jung, Kumar Gaurav, Søren Brunak, Laust Hvas Mortensen, Ewan Birney, Tom Fitzgerald, and Moritz Gerstung. Learning the natural history of human disease with generative transformers. Nature, 647:248–256, 2025.
  31. [33]Alexander Soen, Alexander Mathews, Daniel Grixti-Cheng, and Lexing Xie. UNIPoint: Universally Approximating Point Processes Intensities. Proceedings of the AAAI Conference on Artificial Intelligence, 35(11):9685–9694, 2021.arxiv.org/abs/2007.14082
  32. [34]Yujee Song, Donghyun Lee, Rui Meng, and Won Hwa Kim. Decoupled Marked Temporal Point Process using Neural Ordinary Differential Equations. International Conference on Learning Representations,arxiv.org/abs/2406.06149
  33. [35]Ethan Steinberg, Jason Fries, Yizhe Xu, and Nigam Shah. MOTOR: A Time-to-Event Foundation Model For Structured Medical Records. International Conference on Learning Representations, 2024. Cathie Sudlow, John Gallacher, Naomi Allen, Valerie Beral, Paul Burton, John Danesh, Paul Downey, Paul Elliott, Jane Green, Martin Landray, Bette Liu, Paul Matthews, George Ong, Jill Pell, Alan Silman, Alan Young, Tom Sprosen, Tim Peakman, and Rory Collins. UK Biobank: an open access resource for identifying the causes of a wide range of complex diseases of middle and old age. PLOS Medicine, 12(3):e1001779, 2015.
  34. [36]Chris Whong. FOILing NYC’s Taxi Trip Data. 2014. https://chriswhong.com/open-data/foil_nyc_taxi/.
  35. [37]Michael Wornow, Rahul Thapa, Ethan Steinberg, Jason Fries, and Nigam Shah. EHRSHOT: An EHR Benchmark for Few-Shot Evaluation of Foundation Models. Advances in Neural Information Processing Systems, Datasets and Benchmarks Track, 2023.arxiv.org/abs/2307.02028
  36. [38]Shuai Xiao, Junchi Yan, Xiaokang Yang, Hongyuan Zha, and Stephen M. Chu. Modeling the Intensity Function of Point Process via Recurrent Neural Networks. Proceedings of the AAAI Conference on Artificial Intelligence, 31(1), 2017.arxiv.org/abs/1705.08982
  37. [39]Siqiao Xue, Xiaoming Shi, James Y. Zhang, and Hongyuan Mei. HYPRO: A Hybridly Normalized Probabilistic Model for Long-Horizon Prediction of Event Sequences. Advances in Neural Information Processing Systems, 2022.arxiv.org/abs/2210.01753
  38. [40]Siqiao Xue, Xiaoming Shi, Zhixuan Chu, Yan Wang, Hongyan Hao, Fan Zhou, Caigao Jiang, Chen Pan, James Y. Zhang, Qingsong Wen, Jun Zhu, and Hongyuan Mei. EasyTPP: Towards Open Benchmarking Temporal Point Processes. International Conference on Learning Representations, 2024.arxiv.org/abs/2307.08097
  39. [41]Chenghao Yang, Hongyuan Mei, and Jason Eisner. Transformer Embeddings of Irregularly Spaced Events and Their Participants. International Conference on Learning Representations, 2022.arxiv.org/abs/2201.00044
  40. [42]Qiang Zhang, Aldo Lipani, Omer Kirnap, and Emine Yilmaz. Self-Attentive Hawkes Processes. International Conference on Machine Learning, 2020.arxiv.org/abs/1907.07561
  41. [43]Yuankang Zhao and Matthew M. Engelhard. Balancing Interpretability and Flexibility in Modeling Diagnostic Trajectories with an Embedded Neural Hawkes Process Model. In Machine Learning for Healthcare, volume 298 of Proceedings of Machine Learning Research, 2025.arxiv.org/abs/2504.21795
  42. [44]Ke Zhou, Hongyuan Zha, and Le Song. Learning Triggering Kernels for Multi-dimensional Hawkes Processes. Proceedings of the 30th International Conference on Machine Learning, pages 1301–1309, 2013.
  43. [45]Zihao Zhou and Rose Yu. Automatic Integration for Fast and Interpretable Neural Point Processes. In Learning for Dynamics and Control Conference, volume 211 of Proceedings of Machine Learning Research, 2023.
  44. [46]Simiao Zuo, Haoming Jiang, Zichong Li, Tuo Zhao, and Hongyuan Zha. Transformer Hawkes Process. International Conference on Machine Learning, 2020.arxiv.org/abs/2002.09291

Paper details

vlnc:W_tjsmxbfq
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents