acceptodds
Under review as a conference paper at ICLR 2027

Edge Selection for the Effective use of Piecewise-Constant Distributions as Neural Network Outputs for Event Prediction

Abstract

We study the output representation of a neural network used for next event prediction. We propose partitioning the time axis into a fixed set of intervals and having a neural network output a categorical distribution over them, which we map to a (mostly) piecewise-constant probability density. We present an optimization procedure that selects interval edges in order to maximize data likelihood under the representation. The representation is well suited to processes whose inter-event distribution is a mixture of smooth and sharply peaked components—a pattern we find common in event data recorded from real-world processes.

Primary area:
learning on time series and dynamical systems
Keywords:
point processes, event prediction, neural networks, tpp

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_WmpYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper considers how to represent the next-event time distribution at the output of a neural network. It predicts probability masses on fixed intervals, converts these to a piecewise-constant density, and uses an exponential tail for the last, unbounded interval. The interval boundaries are chosen before network training by dynamic programming on an empirical likelihood objective. The experiments compare output heads on 14 real-world datasets with three backbones, with further analyses of boundary selection and the effect of timestamp precision on likelihood scores.

Critical strengths
  1. I like the decision to keep the backbone fixed when comparing output heads. This makes it much easier to see what the distributional representation contributes.
  2. There is quite a lot of experimental work here: multiple datasets and backbones, repeated runs, confidence intervals, and ablations on interval placement and count. The density-versus-hazard comparison and the experiments with perturbed timestamps and continuous synthetic data are also useful for understanding the results.
  3. The timestamp-precision analysis is, for me, one of the more interesting parts of the paper. Narrow peaks in recorded event times can dominate density-based scores, and the authors do examine this rather than just reporting the gains. The cases where log-normal mixtures do better are useful too.
Critical weaknesses
  1. I am less convinced by the methodological novelty. Kvamme and Borgan [1] already predict a temporal PMF with a neural network and obtain constant-density interpolation by dividing interval mass by interval width; likelihood-based irregular histograms constructed by dynamic programming also have precedents [2], which Section 4 acknowledges. As I understand it, the new part is mainly the particular boundary-selection procedure and its use for history-conditioned prediction, backed by the empirical study. That seems a narrower contribution than the overall construction might suggest, and I would like the paper to be clearer about it.
  2. My main concern is still the size of the likelihood gains. Section 3.1 allows bins as narrow as the local spacing of float32 values. With density equal to mass divided by width, placing these bins around repeated timestamps can make pointwise likelihood much better. On MOOC, for example, NLL goes from 1.02 to -6.23 but MAE only changes from 290.5 to 290.4 (Table 2). Of course, MAE need not improve when a predictive distribution improves. Still, I am not sure what to take from such a large NLL difference here. The jitter experiments help, but I would really like to see an evaluation using interval probabilities, with the same meaningful temporal resolution for every model.
  3. Another question is the use of marginal data to choose boundaries for a conditional prediction problem. The appendix examples with conditional peaks are helpful, but what happens when different histories shift the peaks to different places? I did not get a clear sense of the cost of using one global partition. A fairly simple baseline would help here: an unconditional histogram with the same optimized boundaries. It would tell us how much is gained from fitting the marginal distribution before the history encoder contributes anything.
  4. Appendix C.5 explains the random 6:2:2 splits and the merging of the original Taobao and Amazon splits. The authors also acknowledge the resulting limit on comparison with published results. My remaining question is the splitting unit: are whole sequences assigned to partitions, and can overlapping windows or shared histories appear across them? This needs to be explicit.
Questions for authors
  1. If all models are scored by interval probabilities at a shared temporal resolution, how much of the likelihood advantage remains?
  2. What performance does an unconditional histogram achieve with the same optimized boundaries?
  3. How sensitive is the global partition to history-dependent changes in peak location or distribution shape?
  4. What exactly is assigned to each train, validation, or test split, and how is overlap between windows or histories handled?
4 comments

We thank the reviewer for the valuable feedback provided! We hope to address the points raised.

I am less convinced by the methodological novelty....That seems a narrower contribution than the overall construction might suggest, and I would like the paper to be clearer about it.

We definitely do not intend to overemphasize novelty claims. Actually, our main motivation for writing this paper was to use the proposed method to raise awareness of what we see as important and under-appreciated properties of model-data interaction and issues with how models are being evaluated. The investigation allowed us to present evidence-supported hypotheses for why models score differently (Appendix E). We propose to add the following paragraph to the introduction, which hopefully makes the contribution scope clear and upfront:

As will be described in Section 3.1, our proposed edge selection procedure is one of many which use discretization to relax constraints on distributions for temporal data. Our method uses a known property of maximum-likelihood histograms, that maximizing likelihood over intervals decomposes additively (Jackson et al., 2005), which has been used previously in works such as Simensen et al. (2026) to construct histograms. This work’s procedure for edge selection brings dramatic improvements in likelihood scores to next event prediction on real-world datasets, and we use these results to argue that the presence of concentrations of next-event times in these datasets is tightly linked to and helps explain both our approach's likelihood scores and the scores of previous methods.

Additionally, please have a look at the paragraph added in response to Reviewer vn7Z comment, which extends Section 3.1.

The jitter experiments help, but I would really like to see an evaluation using interval probabilities, with the same meaningful temporal resolution for every model.

Added, as Ablation 5 (Appendix B.5).

Another question is the use of marginal data to choose boundaries for a conditional prediction problem. ...A fairly simple baseline would help here: an unconditional histogram with the same optimized boundaries....

This is a nice control, and we have added it as Ablation 1 (Appendix B.1).

Appendix C.5 explains the random 6:2:2 splits and the merging of the original Taobao and Amazon splits. The authors also acknowledge the resulting limit on comparison with published results. My remaining question is the splitting unit: are whole sequences assigned to partitions, and can overlapping windows or shared histories appear across them? This needs to be explicit.

We have modified the first paragraph of Appendix C.5 (Dataset splits) to clarify this. Including it here for convenience:

For all datasets, each experiment repeat involved randomly sampling training, validation and test sets from the full dataset, targeting a 6:2:2 ratio. For the real-world datasets, which are lists of sequences, this sampling involves randomly allocating whole sequences to train, validation and test splits so that the total event counts in these splits have a ratio close to 6:2:2. These sequences may overlap in time, meaning that the splits do not form a temporal partition. The synthetic sequence generated from the Hawkes process is a single sequence, and train, validation and test sets are formed by partitioning this sequence into three parts.

1. If all models are scored by interval probabilities at a shared temporal resolution, how much of the likelihood advantage remains?

Does the added ablation (ablation 5, Appendix B.5) help answer your question? The experiment's results show a similar pattern to the jitter experiment.

**2. What performance does an unconditional histogram achieve with the same optimized boundaries?

  1. How sensitive is the global partition to history-dependent changes in peak location or distribution shape?**

Does the added ablation (ablation 1, Appendix B.1) help answer your question? The ablation shows that models with and without history encoders have similar performance differences for both the logmix and pc heads, and significant NLL improvements can be attributed to the unconditional distribution with the optimized edges.

4. What exactly is assigned to each train, validation, or test split, and how is overlap between windows or histories handled?

Does the above added paragraph (for Appendix C.5) help answer your question?

Thanks again for the feedback so far; it has been helpful in improving the paper! We are eager and ready to answer any further questions you may have before the rebuttal period ends.

Thanks for running these comparisons. Appendix C.5 answers my question about how the data were split. The no-history result is also useful. On MOOC, removing event history changes the pointwise NLL from −6.23 to −6.14, so most of the gap over the log-normal mixture remains even without history.My main concern was how much weight to put on that large NLL gap. Figure 9 helps answer this. On MOOC, the two heads score nearly the same at interval widths of six seconds and one minute. The piecewise-constant head pulls ahead at finer widths. Bae et al. (ICLR 2023, §5.1) already noted the issue with pointwise NLL, but I found it useful to see how much the gap can change on real data. I am still unsure which widths are meaningful for these prediction tasks and how much of the advantage remains at those widths.The revision makes the empirical contribution clearer. I still see the edge-selection procedure as a fairly modest extension of existing ideas, though, and the new results do not change my overall rating.

Thank you for the continued engagement.

I still see the edge-selection procedure as a fairly modest extension of existing ideas

I don't think we are in strong disagreement here. However, we argue that there is more to this paper than simple score improvements. An achievement of the paper is demonstrating that the ability to concentrate density in narrow intervals is a characteristic that explains the performance differences between existing models (Section 8.2 and Appendix E). We show that the concerns you have also apply to other commonly used heads, especially the logmix head. There are many papers tabulating differences between such models, but a scarcity of explanations for score differences, and we argue that the hypotheses and evidence will be of interest to all people working in the setting of neural TPPs.

Reviewer_7NBNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper studies a piecewise-constant output density for next-event time prediction. A neural network outputs probability mass over fixed nonuniform intervals, with an exponential tail beyond the last finite edge. The interval edges are chosen before model training by dynamic programming to maximize held-out likelihood of an unconditional empirical density. The method is paired with recurrent and transformer stems and compared with log-normal mixtures and several hazard parameterizations on 14 datasets. It reports very large NLL gains on datasets with repeated or highly quantized inter-event times, while MAE changes little. Ablations vary edge count, edge placement, timestamp jitter, density versus hazard, and a synthetic Hawkes process. The experiments are extensive and the paper candidly identifies timestamp precision as the source of much of the gain. That observation, however, undermines the main evaluation: an ordinary continuous density with float32-width bins can assign arbitrarily high density to repeated quantized values. The reported NLL advantage therefore does not establish better probabilistic event prediction.

Critical strengths
  1. The edge-selection objective is simple and the dynamic program is explained clearly, including candidate construction and complexity. The output head is easy to implement, normalized, and compatible with different sequence encoders.
  2. The empirical scope is unusually broad: 14 datasets, three stems, several output heads, ten repeats, and targeted ablations. The timestamp-jitter and Hawkes experiments are especially valuable because they reveal that the strongest likelihood gains depend on narrow repeated-time peaks, and the limitations section openly questions density-based evaluation at arbitrary precision.
Critical weaknesses
  1. The central likelihood comparison uses the wrong observation measure for the data it emphasizes. Repeated, rounded timestamps contain discrete atoms, yet every model is scored as a Lebesgue density. The proposed edge search permits widths down to a float32 unit in the last place, so density and test log likelihood can grow as the storage grid becomes finer without improving probability assigned to any meaningful time interval. The extreme negative NLL values, dependence on timestamp inequality, and degradation under jitter are evidence of this metric artifact, not validation of a continuous-time predictor.
  2. The protocol further weakens the empirical claim. Published splits for Amazon and Taobao are merged and randomly resampled, preventing comparison with prior work and potentially sharing user- or entity-specific regularities across partitions. The significance rule treats ten random splits as independent standard errors, uses no paired test or multiple-comparison correction, and follows manually inspected learning-rate overrides. MAE is usually unchanged or worse, and log-normal mixtures win on the Hawkes process, so the claimed benefit does not generalize beyond quantized benchmark timestamps.
Questions for authors
  1. What remains of the advantage when the observation model uses a proper mixed discrete-continuous measure, or when likelihood is computed for probability mass over a prespecified measurement resolution? Please compare against uniform dequantization and resolution-aware discrete baselines without float32-ULP bins.
  2. Can you rerun the main comparison on the official fixed splits with entity-disjoint partitions where applicable, paired statistical tests, and interval-based calibration scores? A sweep over declared timestamp resolutions should show whether edge optimization adds value beyond memorizing repeated grid points.
3 comments

We thank the reviewer for their careful reading and valuable feedback, and apologies in the delayed response. Each point below led to a concrete change to the paper. We will address Weakness 1 in this reply, and Weakness 2 and the questions subsequently.

The central likelihood comparison uses the wrong observation measure for the data it emphasizes...

We agree that there is a mismatch between properties of the event data and the continuous-density evaluation and that our model's strong scores can be attributed to this mismatch.

Actually, one of the main reasons we wrote this paper was to raise awareness of this mismatch. Our datasets and evaluation proceedure are standard when considering the related work in Section 5.1. We hope to present field of neural temporal point processes with the contrast:

  • if continuous density is maintained as the core evaluation metric for sequences recorded from real-world processes, we show that output heads that can concentrate density can dominate. This is not unique to our method: we show that the existing logmix and nn output heads can do it (Figure 4f, and Appendix E).

or

  • if such concentrations are not meaningful for the intended task, the prediction task should be revised. We outline alternatives such as the next-event prediction task being changed to include a resolution of interest (Section 9, L488-500) with probability mass evaluation.

We think this motivation for the paper was not made clear.

This feedback point has led to the last paragraph of the introduction being extended to emphasize this goal more clearly, and how we provide empricical evidence supporting a reconsideration of established evaluation practices. Probability mass evaluation over intervals has also been added as Ablation 5.

Published splits for Amazon and Taobao are merged and randomly resampled, preventing comparison with prior work and potentially sharing user- or entity-specific regularities across partitions.

We split both datasets in the same way the EasyTPP benchmark suite does: the data is a list of sequences, one sequence per user, and train/val/test sets are formed by randomly assigning sequences to a partition. Thus, users are not shared across splits. The only difference is that we consider multiple splits so as to have a better measure of variability due to data splits.

The reviewer is correct to point this out however, as it is not explained clearly in the text. This remark led to an improved description of the dataset splits in Appendix C.5.

The significance rule treats ten random splits as independent standard errors, uses no paired test or multiple-comparison correction

This is a valid and important criticism. We had used independent standard errrors for simplicity. We agree with the reviewer that this is distinctly less informative than paired tests.

We appreciate this feedback point and we have taken it on board by changing all tables that use confidence tests to now use paired tests with Holm-Bonferroni correction. To keep table captions at a reasonable length, we have added a section (Appendix A.1) that describes the proceedure in detail, to which tables now reference.

follows manually inspected learning-rate overrides

We may not have been clear about how learning rate selection was carried out.

We make sure all models have a fair comparison by using learning rate sweeps to find appropriate learning rates for each dataset-model pair. The proceedure was largely automated so that it can be easily replicated by others. From the output of the automated process, we manually inspect every sweep and learning rate suggestion to ensure they are appropriate. There was a small number of cases where unstable training at high learning rates produced complicated traces that the automatic proceedure did not interpret correctly. In these cases, we step in to manually select a suitable rate, "overriding" the automatically chosen one.

This remark led to Appendix F getting a more detailed description of the learning rate selection procedure.

MAE is usually unchanged or worse

The rnn-pc model (ours) has the lowest (best) mean rank in terms of MAE across the 14 datasets. This is shown in Table 7. We do, however, agree with the reviewer in that the MAE differences are small.

log-normal mixtures win on the Hawkes process, so the claimed benefit does not generalize beyond quantized benchmark timestamps.

We agree that the pc head comes second to the logmix head on the synthetic hawkes sequence. Both approaches' scores are close when considering their differences to the other heads.

We do not think this one result undermines our findings. The paper is focused on data recorded from real-world processes, and the synthetic Hawkes sequence is used to show that our proposed method still works for data that is known to be continuous.

Regarding "beyond quantized benchmark timestamps", we may have failed to point out other sources of narrow density peaks in the real-world datasets. For example, the sequences of badge reward times on Stack Overflow is very smooth: data is seconds with the maximum inter-event time being 15 years (75th quantile is 155 days). And yet, there is a very narrow (multiple minutes) concentration of inter-event times at around 365-days, presumably related to badges tied to yearly events (Figure 13). As the concentration extends over multiple minutes while the quantization is seconds, the effect is not from timestamp rounding.

This remark led to Section 8.1 being extended to highlight that the peak in the Stack Overflow dataset is an example of a narrow peak that is not related to timestamp quantization.

What remains of the advantage when the observation model uses a proper mixed discrete-continuous measure, or when likelihood is computed for probability mass over a prespecified measurement resolution? Please compare against uniform dequantization and resolution-aware discrete baselines without float32-ULP bins.

This comment led to the addition of Ablation 5 (Appendix B.5) where probability mass over a range of fixed intervals is evaluated. Results are similar to the jitter experiment.

We are here further discussion, and are happy to keep improving the paper with more feedback.

Thanks again for the feedback so far; it has been helpful in improving the paper! We are eager and ready to answer any further questions you may have before the rebuttal period ends.

Reviewer_vn7ZOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose to model the time to event (TTE) density as a piecewise constant density function. They consider equal-width intervals, equal-quantile intervals, and variable-width intervals learned by maximizing the marginal likelihood of observed event times via dynamic programming. Results focus on event sequence datasets (predict next event density given event history) rather than TTE/survival analysis datasets (predict event density given covariates). Results show benefit across many datasets compared to a wide variety of baselines.

Critical strengths

The paper is well-written and enjoyable to read, which is increasingly uncommon and I appreciate it very much. The authors explain the approach clearly and do a nice job citing relevant work, though with important caveats discussed in the next section. Results are strong and clearly illustrate the benefit of the approach compared to the baselines the authors considered. I have not seen the interval selection approach proposed here in other TTE work, though I am not certain it has not been.

Critical weaknesses

The use of piecewise constant densities in TTE prediction has a long history going back to multi-task logistic regression [1]. There have been countless variations on this, with DeepHit [2] being among the most well-known. Both linear and neural MTLR are implemented in PySurvival [3]. Approaches to learn interval edges are not common but have been explored [4].

In light of this, the contribution here is (a) the edge selection procedure, which seems to be adapted from the literature on variable-width histograms but is new in the TTE prediction setting to my knowledge; and (b) strong empirical results supporting the overall approach. This is incremental though still valuable in my view.

A minor weakness is that the approach does not support learning from right-censored outcomes. In general this is important in TTE settings, though less so here given the authors' focus on inter-arrival times in event sequences.

[1] Yu CN, Greiner R, Lin HC, Baracos V. Learning patient-specific cancer survival distributions as a sequence of dependent regressors. Advances in neural information processing systems. 2011;24.

[2] Lee C, Zame W, Yoon J, Van Der Schaar M. Deephit: A deep learning approach to survival analysis with competing risks. InProceedings of the AAAI conference on artificial intelligence 2018 Apr 26 (Vol. 32, No. 1).

[3] https://square.github.io/pysurvival/

[4] Hickey J, Henao R, Wojdyla D, Pencina M, Engelhard M. Adaptive discretization for event prediction (adept). InInternational Conference on Artificial Intelligence and Statistics 2024 Apr 18 (pp. 1351-1359). PMLR.

Questions for authors

I think there is quite a bit of relevant literature and context to be added, as described above, but nevertheless there is a nice contribution here.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.