acceptodds
Under review as a conference paper at ICLR 2027

Smooth Neural Point Processes via B-Splines

Abstract

Temporal point processes (TPPs) provide a general and flexible framework for modeling sequences of events in continuous time. Neural networks have been successfully employed to model TPPs in a highly expressive and data-driven way. Neural TPPs are typically trained via Maximum Likelihood Estimation (MLE) by minimizing the negative log-likelihood (NLL), which depends on both the conditional intensity function (CIF) and its integral over time, the compensator. Recent neural TPP approaches enable exact evaluation of the NLL without numerical integration. However, these methods typically model the compensator rather than the CIF directly, impose constraints on the neural network architecture, and are computationally expensive during training, as event contributions to the NLL are evaluated sequentially rather than in parallel. In this work, we propose a novel neural TPP model that directly parametrizes the CIF as a non-negative combination of B-spline basis functions, whose coefficients are predicted by a neural network. This formulation enables exact evaluation of the NLL, preserves full flexibility in the neural architecture, allows efficient parallelization during training, and naturally supports CIF smoothness regularization by penalizing the integral of the squared second derivative. Experiments on both synthetic and real-world datasets show improved computational efficiency and predictive accuracy compared to modern neural TPP baselines.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Temporal point processes, Neural temporal point processes, B-splines

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_8ktNOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SmoothTPP, a neural temporal point process that models the conditional intensity function using non-negative B-spline coefficients predicted from event history. This approach enables analytical likelihood evaluation, flexible neural architectures, and efficient smoothness regularization. Experimental results demonstrate the effectiveness of the SmoothTPP.

Critical strengths

The B-spline parameterization provides a simple and tractable way to model intensities with analytical integration.

Critical weaknesses
  1. The spline construction is not fully specified, particularly the boundary knots and intensity behavior beyond the training range. Besides, the inter-arrival likelihood omits the terminal survival term. These details are important for reproducing the method and understanding the exact-likelihood formulation.

  2. The experiments do not report variability across random seeds and omit comparisons with additional relevant TPP methods, such as ADD-THIN. The multivariate extension is not evaluated experimentally.

Questions for authors
  1. How are spline boundary knots selected, and how are terminal censored intervals incorporated into the likelihood?

  2. Event sequence modeling involves dependencies on past events. How does SmoothTPP implement parallel training while ensuring that each event’s intensity is conditioned only on its preceding history?

  3. Do all models use the same conditional-median prediction protocol for MAE evaluation?

1 comment
AuthorsOpenReview ↗

Response to Reviewer 8ktN

We thank the reviewer for the comments.

Responses to the comments:

  1. We did not specify the behavior outside the spline domain because, in our experiments, no observed test inter-arrival times exceeded the maximum training inter-arrival time. Nevertheless, this may occur, so an explicit tail definition is necessary. Several solutions are possible: a simple choice is to keep the intensity constant beyond the spline domain, at its boundary value. Our boundary-knot construction is compatible with this extension: since we use cubic splines, each boundary knot has multiplicity four, ensuring that the intensity at the right endpoint equals the last spline coefficient.

    Regarding the survival term, this can be incorporated into our model without any difficulty. However, to our knowledge, this is also omitted in EasyTPP, the framework used for our benchmarks. Moreover, the datasets used in the experiments provide only the observed event sequences, without information about the corresponding observation window.

  2. We will add experiments with multiple random seeds, multivariate datasets, and ADD-THIN as an additional baseline.

Answers to the questions:

  1. The boundary knots are currently placed at the endpoints of the domain, from zero to the maximum inter-arrival time observed in the training set. Since we use cubic splines in our experiments, we assign multiplicity four to each boundary knot.

    As the datasets used provide only event sequences and not the total observation time, terminal censored intervals are currently not included, as is also the case in EasyTPP. It is also reasonable to expect these contributions to be negligible.

  2. In SmoothTPP with a Transformer architecture, this is ensured by causal masked attention, which prevents each prediction from attending to future events.

  3. Yes, all models use the conditional median for the MAE comparison. In SmoothTPP, it is computed analytically, whereas for the other models it is estimated empirically from samples generated via thinning.

Reviewer_4iskOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposed a novel model named SmoothTPP, where the conditional intensity of inter-arrival times is modelled by B-spline functions and the coefficients are neural network outputs. This approach has the advantage of having a closed-form compensator compared to existing intensity-based neural marked temporal point processes. Experiments use a transformer as an event encoder and parameterize the conditional intensity function with a cubic B-spline and 20 internal knots, evaluated on test (negative) log-likelihood and mean absolute error on 7 simulated datasets and 2 real-world datasets.

Critical strengths

The approach of applying B-splines to model the conditional intensity function is both novel and well-motivated to have closed-form compensators. The paper is also well organized, clearly presented, and makes a great effort at reproducibility.

Critical weaknesses

Despite its novel and interesting idea, the most critical weakness is that this paper contains many factual errors, from math to references. To list a few examples indicating unfamiliarity with the framework and literature:

Math:

  • From line 105 to line 112, the log-likelihood equation silently drops the survival term (the integral from TNT_N to TT between the last observed event and the right end of the observation window). This results in two issues: (i) implicitly assuming that the observation ends at an event, and (ii) artificially improving the log-likelihood without this term if a sequence does not end with an event.
  • Line 114 only defines ϕ(τi)=λ(Ti)\phi(\tau_i) = \lambda(T_i), whereas line 115 uses ϕ(s)ds\phi(s) ds that is not defined.
  • Even if we can generalize line 114 to any point in time and regardless of the survival term, ϕ(τi)=λ(Ti)\phi(\tau_i) = \lambda(T_i) itself is an incorrect transformation for the change-of-variable derivation.

References:

  • Line 31 (and line 466): The point process textbook has only two authors, Daley and Vere-Jones (without "et al.").
  • Line 48 (and line 486): Mei is the second author of AttNHP rather than the first author.
  • Line 49 (and line 460): S2P2 was published at NeurIPS 2025, not 2026.
  • Minor ones: line 485 would be great to capitalize "Hawkes", and line 514 would be great to cite the ICLR 2024 version of the EasyTPP paper rather than the arXiv link.

Additionally, there are several inconsistencies in the experimental setup and implementation; clarifying them will greatly improve the trustworthiness of the experimental results. See more concrete questions in "Questions For Authors".

Questions for authors

It would be great if the authors could ensure the correctness of their derivation, parameterization, and evaluation to explore this idea further. Below are questions to help achieve this. Without addressing these questions, the current results are not reliable enough to draw any conclusion. Elaborating on more technical details would be strongly encouraged in any future manuscript.

Questions:

  • As discussed in the weaknesses above, can the authors explicitly write down the change of variables and the log-likelihood derivation from lines 105 to 112?
  • In line 75, can T1=0T_1 = 0?
  • If the log-likelihood misses the survival term for SmoothTPP, then it is not comparable with 3 out of 4 baselines (NHP, AttNHP, S2P2) that have the survival term in their log-likelihood if the sequences do not always end with an event. Is there any experimental setup to incorporate this and make the negative log-likelihood (NLL) results in Table 1 comparable?
  • For the next event prediction, if baseline models use the thinning algorithm, the hyperparameters should be specified to ensure they do not bias the results at their best effort. From the uploaded config file, only 21 samples were used for thinning. It is recommended to include variances over multiple random seeds in both Tables 1 and 2 for any statistically significant conclusions (e.g., "consistently improved..." (line 417) and "remains competitive..." (line 419)).
  • Figures 1 and 2 only plot the learned conditional intensity function (CIF) of SmoothTPP against the ground truth. For simulated datasets, plotting the predicted CIF for all models (including both SmoothTPP and baseline models) against the ground truth would be suggested, which will greatly help strengthen the results and back up the summarized quantitative results in Table 1.
  • The paper states an 80%/20% train/test split in line 253 for each sequence, while line 303 mentions a validation set obtained from 20% of the training data is very confusing. To clarify, does this imply a 64%/16%/20% train/validation/test split for each event sequence?
  • Similarly, for real-world datasets: (1) is there any reason why two datasets have different ways of train/test splits: (i) within a sequence or (ii) across sequences? Is there a subset for validation? Is there a reason why only 100 and 50 sequences are evaluated for each dataset, and would this affect the variance of results across several random seeds? What are the average lengths and processing procedures for these datasets?
  • Regarding the validation set and hyperparameter selection, is there a reason for the specific architectures chosen for these baselines, e.g., by roughly matching the parameter counts? If not, is there a hyperparameter sweep on the validation set for each baseline model and per dataset to give them the best chance?
  • Lines 292-293 are confusing: if "The CIF curve over future times is predicted by taking as input the inter-arrival times of the last 20 events", how is the CIF curve evaluated in Figures 1 & 2? Does this relate to the experimental setup of the next event prediction? Because the supplementary code SmoothTPP.ipynb also has some "cap at T" in predicting the next inter-arrival time median, could the authors discuss any potential implications on the MAE estimation?
  • What is the impact of different inter-arrival time distributions for different datasets on SmoothTPP, as the knots are distributed at quantiles? For example, the Retweet dataset from EasyTPP has long-tailed inter-arrival times; will this affect SmoothTPP's performance?
  • Is SmoothTPP's performance sensitive to equidistant knots, knots placed at quantiles, or the number of knots? Such investigations would be very helpful for a comprehensive understanding of the proposed models in addition to leaderboard comparisons.
  • Is there any limitation of the B-splines approach? Explicitly pointing this out will help practitioners choose different models for specific applications.
1 comment
AuthorsOpenReview ↗

Response to Reviewer 4isk

We thank the reviewer for the very detailed review and for pointing out some typos, which we will correct.

Answers to the questions:

1) We thank the reviewer for identifying an imprecision in our notation. The history-dependent intensity should be indexed by the inter-event interval. Setting T0=0T_0=0 and τi=Ti−Ti−1\tau_i=T_i-T_{i-1}, we define ϕi(u)=λ(Ti−1+u)\phi_i(u)=\lambda(T_{i-1}+u) for 0<u≤τi0<u\leq\tau_i, with the conditional intensity evaluated along the observed history (before the event at the right endpoint). The substitution s=Ti−1+us=T_{i-1}+u, with ds=duds=du, gives

=\int_0^{\tau_i}\phi_i(u)\mathrm{d}u =\Phi_i(\tau_i).$$ Moreover, $\lambda(T_i)=\phi_i(\tau_i)$. For the terminal interval in an observation window $[0,C]$, define $\phi_{N+1}(u)=\lambda(T_N+u)$ for $0<u\leq C-T_N$, conditioned on the history through $T_N$ and no subsequent events, and $\Phi_{N+1}(v)=\int_0^v\phi_{N+1}(u)\mathrm{d}u$. The complete log-likelihood is $$\ell=\sum_{i=1}^{N}\left[\log\phi_i(\tau_i)-\Phi_i(\tau_i)\right] -\int_0^{C-T_N}\phi_{N+1}(u)\mathrm{d}u =\sum_{i=1}^{N}\left[\log\phi_i(\tau_i)-\Phi_i(\tau_i)\right]-\Phi_{N+1}(C-T_N).$$ **2)** Under the paper's convention, no: $0$ marks the start of observation, and $T_1>0$ is the first subsequent event. For a process with an ordinary intensity with respect to time, an event at the deterministic time zero has probability zero. **3)** The terminal survival term can be incorporated into SmoothTPP's log-likelihood without difficulty, as done in 1). It was omitted in the paper submission because the datasets used provide event times but not the total observation duration, so this contribution could not be evaluated. To our knowledge, this term is also omitted in EasyTPP, and we followed the same convention for consistency with the benchmarks. **4)** We will follow this suggestion by repeating experiments with multiple random seeds and analyzing the number of samples required for convergence of the empirical median for baselines using thinning. **5)** Thank you for the suggestion. We will add plots comparing all models' predicted CIFs with the ground truth. **6)** Yes, exactly. Of the 80% initially allocated to training, 20% is reserved for validation, yielding an overall **64% / 16% / 20%** train/validation/test split. **7)** The different train/test splits and the selection of the top 100 sequences follow the protocol of Omi et al., which we initially chose as the main reference for our comparisons; the other models were added later. A validation set is also used, consisting of 20% of the training sequences. We will add statistics on average sequence lengths. **8)** Yes. Hyperparameters controlling model size were selected to obtain comparable parameter counts across models. For the remaining hyperparameters, we retained the EasyTPP defaults. **9)** Figures 1 and 2 concatenate curves from successive predictions. A new prediction is made whenever a new event occurs, based on the most recent past 20 events. The cap at $T$ in the code concerns a separate issue: behavior outside the spline domain is currently undefined. This did not affect the reported results: all test inter-arrival times were below $T$, the maximum training inter-arrival time, and all medians obtained by bisection for MAE evaluation were also below $T$. Several strategies can handle inter-arrival times above $T$. The simplest approach we recommend is to extend the intensity beyond the spline domain as a constant equal to its last valid value. **10)** Thank you for this suggestion. We will add Retweet to the real-world datasets to address these questions empirically. **11)** We have observed greater sensitivity to knot placement (quantile versus equidistant), particularly for distributions with sharp initial peaks followed by rapid decay. Sensitivity to the number of knots is lower, provided there are enough knots to represent the underlying conditional intensity adequately. Oscillations associated with an excessive number of knots are controlled through the regularization parameter $\lambda$. **12)** In our view, the main limitations are the additional regularization parameter $\lambda$ to tune and, for datasets with many event types, the need for a sufficiently large network to produce an output whose dimension scales as the number of event types times the number of spline coefficients. These are trade-offs for greater tractability, full flexibility in the choice of neural architecture, faster training through exact log-likelihood evaluation, and efficient inference through obtaining the entire future intensity curve in a single forward pass.
Reviewer_MKUiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SmoothTPP models the history-conditioned intensity of a temporal point process with B-splines. A Transformer encodes the event history and predicts nonnegative spline coefficients. The basis functions can be integrated analytically, allowing exact likelihood evaluation for the specified intensity, and a roughness penalty discourages excessive curvature. Experiments on seven synthetic and two real-world datasets report lower MAE and an approximately twelvefold speedup from a parallel training formulation.

Critical strengths
  1. The spline head is simple, which I see as a strength. Nonnegative coefficients and basis functions give a nonnegative intensity within the spline domain, and the integral can be computed analytically without Monte Carlo error.
  2. I also like the explicit roughness penalty. It gives a reasonably transparent way to control fluctuations, and the precomputed matrix makes the quadratic penalty cheap to evaluate.
  3. The MAE results look promising, although I have a concern about the prediction rule below. The likelihood gain on Meme is also worth following up. There is enough detail in the architecture and training description to give a useful starting point for reproduction.
Critical weaknesses
  1. I would have liked to see a discussion of Deep Random Splines for Point Process Intensity Estimation of Neural Population Data [1]. It also generates nonnegative spline intensities with neural networks and integrates them analytically for likelihood evaluation conditional on those intensities. Its main construction uses alternating projections, but Section 3.2 also reports trying B-splines with nonnegative coefficients. SmoothTPP instead predicts a next-event intensity from observed history and adds a roughness penalty, whereas [1] uses a latent-variable model for neural population data. These are meaningful differences, so I would not regard the methods as equivalent. Still, this prior work is relevant to the claimed novelty of the spline construction and should be discussed.
  2. Why does monotonicity imply sequential computation? I do not see that connection. With the history representations already computed, the cumulative intensity and its derivatives can be evaluated in batches. Perhaps the intended point is about the encoder or a particular implementation, but that would be a different explanation for the bottleneck. As written, the argument does not convince me that changing the output parameterization is what enables parallelism.
  3. The twelvefold speedup initially sounds like a comparison with other models, but it is between two training formulations of SmoothTPP. Even within that comparison, sequence length and the organization of prediction targets change. I am not sure how much of the gain should therefore be attributed to the proposed formulation. Matching effective context, target-event counts, and loss normalization would make this easier to assess. I would also like to know the time to reach comparable validation performance, alongside throughput; the current result alone says little about efficiency relative to the baselines.
  4. The MAE comparison is a more immediate concern for me. SmoothTPP predicts the conditional median, while the original EasyTPP protocol [2] uses the conditional mean. Were all baselines changed to use medians? I could not tell. Since the median minimizes expected absolute error, a change in point estimator can improve MAE even when the predictive distribution stays exactly the same. This does not mean the reported gains are spurious, but I would not attribute them to better distribution modeling until the prediction rules are clear.
  5. What happens if only the output head is changed and the history encoder stays fixed? That is the comparison I miss most. For regularization, Section 4.2 does include alpha = 0 in the parameter search, and Figures 1 and 2 show its effect on the intensity curves. What I would still like is a quantitative NLL/MAE comparison with and without the penalty, rather than relying on the selected alpha and the plots. Variability across runs would also help interpret the smaller likelihood differences.
  6. I could not work out how the intensity is defined beyond the spline domain. This matters for the prediction rule, not just for implementation. If the total integrated intensity is finite, there is positive probability of no subsequent event. That can be a valid modeling choice, but what does it mean here, and what does the model return when the cumulative intensity never reaches the threshold for a finite median? If the authors instead extrapolate the spline polynomial, they need to explain why the extrapolated intensity remains nonnegative. I would like the tail construction and the median calculation spelled out together.
Questions for authors
  1. Relative to the neural spline intensity model in [1], what benefit comes from the history-conditioned construction and the roughness penalty?
  2. Were conditional medians used for every baseline in the MAE comparison?
  3. How do the training times compare after matching context length, target events, and loss normalization?
  4. With a fixed history encoder, how much does the spline head improve the results, and what further change comes from regularization?
  5. How is the intensity defined beyond the spline domain, and what prediction is returned when no finite median exists?
1 comment
AuthorsOpenReview ↗

Response to Reviewer MKUi

We thank the reviewer for the insightful comments, which we believe will help substantially improve our contribution.

Responses to the comments:

  1. Thank you for bringing this work to our attention. We will discuss it in the introduction as relevant prior work.

  2. We were referring specifically to FullyNN as introduced in the original paper, which was inherently sequential. However, the FullyNN implementation in EasyTPP already supports batched computation, as the reviewer correctly points out. Our statement was therefore too strong, and we will remove it.

    We would nevertheless like to highlight an additional computational advantage of our approach: a single forward pass provides the intensity curve for all future times, rather than its evaluation at just one specific time. To the best of our knowledge, our model is currently the only one offering this property without assuming a particular functional form for the future intensity.

  3. We agree that the computational efficiency comparison could be improved, and we will work on this, potentially by also implementing SmoothTPP within EasyTPP and conducting all comparisons using the same library, datasets, context windows, and experimental setup.

  4. The comparison was indeed performed using medians. As the reviewer correctly observes, median estimation is not natively supported by EasyTPP, so we implemented it ourselves. Specifically, following a protocol similar to the one EasyTPP uses for the mean, we generated N = 21 samples and computed their empirical median. For SmoothTPP, the median was instead computed analytically. For sufficiently large N, the empirical median should converge to the corresponding distribution's median. To verify whether N = 21 is sufficient, we could conduct an ablation study with larger sample sizes.

  5. Isolating the contribution of the spline parameterization alone is not straightforward, particularly when using a Transformer. The closest comparison would probably be a FullyNN implementation with a Transformer encoder. However, to our knowledge, this is not implemented in EasyTPP, and the networks would still not be entirely equivalent, since FullyNN would include an additional feed-forward network component.

    Regarding regularization, we will add quantitative metrics. We can already report that the results are fairly robust to the choice of the regularization parameter lambda: its effect on oscillations is more apparent in the intensity plots than in the NLL/MAE metrics.

  6. In the current implementation, we did not specify the behavior outside the spline domain because the test set never contained inter-arrival times exceeding those in the training set and thus falling outside that domain. Nevertheless, the concern raised by the reviewer is valid. As a simple choice, we propose extending the intensity beyond the spline domain as a constant equal to its last valid value, analogous to a background intensity in Hawkes processes, to which the intensity returns once the self-exciting effects of previous events have subsided.

Answers to the questions:

  1. In the paper mentioned by the reviewer, the entire history is summarized in a single latent vector. We believe that this representation may lose substantial information for long histories. Furthermore, that paper assumes inhomogeneous Poisson process dynamics, whereas our model is more flexible and makes no specific assumption about the process type.

  2. Yes, as explained in point 4 above, empirical medians were used for the baselines.

  3. We will update this answer once the experiments are complete.

  4. As explained in point 5, isolating the contribution of the parameterization alone is not straightforward, and we welcome suggestions. In any case, we do not necessarily expect the spline representation itself to improve predictive performance. The main advantage is obtaining the entire future conditional intensity curve in a single forward pass, rather than evaluating it at a single time as in FullyNN. This offers benefits for inference and prediction, whether using thinning or analytical median calculation.

  5. As explained in point 6, the simplest choice is to assume a constant intensity beyond the spline domain, equal to its last available value. With a strictly positive tail intensity, the median is always finite and can be computed analytically in SmoothTPP.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.