acceptodds
Under review as a conference paper at ICLR 2027

Sharing Temporal Structure for Joint Time-Mark Prediction

Abstract

Distinct event types can exhibit different timing profiles without requiring separate temporal components. We introduce Shared Temporal Mixture (STMix), a history-conditioned joint time-mark model that uses shared temporal structure for both learning and probability queries. Each mixture component pairs a temporal density with a categorical mark distribution, allowing each type to combine the same temporal components with different weights. During joint likelihood training, observed times and marks determine component responsibilities, enabling events of different types to shape common temporal patterns. Tractable component cumulative distributions express marked next-event window probabilities and forecasts after an event-free wait as finite sums without numerical time integration. A query algorithm prepares components once per observed history and reuses endpoint evaluations across types, horizons, and requests. Experiments on four real-world datasets demonstrate competitive joint prediction; matched structural comparisons on three datasets achieve better test likelihood with 69–86% fewer total parameters than type-specific models. Query experiments show competitive window forecasts, while endpoint reuse yields 1.38–2.60x speedups over a vectorized implementation of the same model without caching.

Primary area:
learning on time series and dynamical systems
Keywords:
marked temporal point processes, joint time-mark modeling, shared temporal components, mixture models, probabilistic forecasting

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

This paper focuses on conditional generation and next event prediction: conditioned on ii events and predicting the time and mark of the i+1i+1th event. In addition to the point estimate, the proposed model STMix can also efficiently predict the probability of the immediate next event during some observation window by reusing the survival cache. In experiments, STMix is compared with four competitive baseline models for both next event prediction and window predictions in terms of quality and efficiency.

Critical strengths

The visualization is very good, such as Figure 1, which greatly helps explain the novelty of the task and approach.

Critical weaknesses

This paper contains incorrect knowledge about temporal point processes, and it does not demonstrate a good understanding of the existing literature. For example, a fundamental assumption of temporal point processes (TPPs) is that there are no concurrent (or simultaneous) events, i.e., the delta time between two consecutive events is strictly positive [1]. This paper predicts the next gap T=ti+1−tiT = t_{i+1} - t_i where TT can be positive or zero, which could have practical values, but it does not fall into the TPP definition/framework.

Additionally, the log-likelihood is defined in Eq. 5, where the survival term is missing for an observation window for a sequence log-likelihood (i.e., no event after the last observed event during the observation window). While this objective has been used in the literature and is especially common in density-based parameterization [2,3], this paper takes a CDF-based approach and compares it with other models where their sequence log-likelihood is defined to include the survival term (AttNHP, S2P2). A lot of experimental details are missing to interpret the results and ensure a fair comparison.

[1] Daley, D. J., & Vere-Jones, D. (2003). An introduction to the theory of point processes: volume I: elementary theory and methods. New York, NY: Springer New York.

[2] Intensity-free learning of temporal point processes, ICLR 2020.

[3] Transformers for Mixed-type Event Sequences, NeurIPS 2025.

Questions for authors

Following up on the weaknesses above, could the authors help elaborate on these questions:

  • Many intensity-based marked temporal point processes model the joint distribution of event time and mark. What is the main difference from the proposed model (in addition to the CDF parameterization)?
  • How is the event encoder hh initialized before any event? Can the STMix model also perform unconditional generation (i.e., predict the first event)?
  • Where do the numbers in Figure 2 come from? For example, is there any hyperparameter sweep performed for each model before comparing the number of parameters and test NLL?
  • Are the NLL comparisons fair for different baseline models? For intractable integrals of intensity-based models (AttNHP, S2P2), how was the NLL estimated?
  • In the window prediction setup: how is the window log loss defined, and how is the window chosen for the experiments? I saw Eq. 22 (on page 24) in the appendix that defines the window log loss, but it would be good to formally define the problem in mathematical terms and equations, and explicitly document the experimental setup and metrics definition in the main text.
  • How does window prediction work for the baseline models (Figures 3 & 4)? It is unclear how these models are adapted for the novel task and how they are evaluated.
Reviewer_gvB3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes STMix, a joint model of the next event time and type. The main idea is to use a small set of shared temporal components for all event types, while each type combines these components with different weights. This reduces parameters and makes time-window probability queries efficient. Training is standard NLL.

Critical strengths
  • Simple and intuitive way to share temporal structure across event types
  • Large parameter reduction compared to having a single time mixture per event type: p(time | mark) p(mark)
  • Efficient and exact window probabilities and conditioning after waiting without an event
  • Experiments and ablations are good
Critical weaknesses
  • Limited novelty. The model is essentially a standard mixture of products with shared temporal components
  • Marks improving temporal learning mechanism is perhaps a bit overstated. This is not a new training method as it follows directly from standard NLL training of the proposed joint mixture
  • The key assumption is restrictive, which is also a selling point
  • The number of shared components is fixed to 4 which seems like a small number
  • Results are mixed
  • The query-speed contribution largely comes from reusing cached endpoint evaluations
Questions for authors

See weaknesses.

Reviewer_2GL2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes STMix, a history-conditioned joint distribution for next-event time and mark. Four shared log-time spline components are combined with categorical mark distributions, so each mark obtains a distinct timing profile through component weights while all marks contribute to the same temporal component bank. The product-mixture form yields analytic mark marginals, marked window probabilities, and survival-conditioned forecasts after an event-free wait. A query implementation decodes the history once and caches component CDF and survival values at repeated endpoints. The paper also studies whether mark-loss gradients through density-posterior responsibilities improve temporal learning. Across four datasets, STMix is competitive but loses joint NLL to S2P2 on three. A capacity-unmatched type-specific comparison favors sharing, while a matched-parameter comparison gives mixed results. Window scores are often competitive, and endpoint caching accelerates a constructed MOOC query workload. The model is carefully presented, but the current evidence is insufficient to support the broad claims about sharing and practical query efficiency.

Critical strengths
  1. STMix is a coherent normalized joint density. Its component interpretation connects time-mark dependence, direct mark feedback, and probability queries in one representation. The formulas for zero gaps, finite windows, and event-free waiting are explicit, and the survival reweighting is exact for the immediate next event.
  2. The paper makes a serious effort to isolate mechanisms. It includes posterior-stop, fixed-gate, learned-gate, independent, type-specific, and parameter-budget controls; reports decomposed time and mark losses; uses proper window scores; and compares cached and uncached implementations of the same fixed model to verify that caching does not alter predictions.
Critical weaknesses
  1. The main sharing comparison is strongly capacity-unmatched: STMix has four temporal components total, whereas the control has four per type and 69 to 86 percent more parameters. The matched-budget experiment is run on only Amazon and MOOC and reverses direction on Amazon, where the compact type-specific model has better test NLL. Different run groups also yield substantially different STMix references on MOOC. A component-count and low-rank sweep across all datasets is needed before attributing generalization gains to sharing rather than regularization or optimization.
  2. The practical and novelty claims are limited. Product mixtures of continuous kernels and categorical marks are established, and the query contribution is largely endpoint memoization. The largest caching result uses one MOOC validation workload, excludes history encoding, and compares against an uncached implementation of the same model. In end-to-end comparisons, LNM is faster on Amazon and the learned gate has better MOOC window scores. STMix also trails S2P2 joint NLL on three of four datasets, while uncertainty is based on only three training seeds.
Questions for authors
  1. Can you repeat the matched-parameter comparison on all four datasets and sweep the shared component count, components per type, and factor rank under paired initializations? Reporting paired confidence intervals would clarify when sharing helps beyond the regularization induced by a much smaller head.
  2. Can you benchmark realistic held-out query traces that include history encoding, cache construction, memory use, and cache invalidation, and compare with an optimized analytic type-specific mixture using the same caching strategy? Please also evaluate post-wait forecasts on test rather than checkpoint-selection data.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Follow-up on our responses

Dear Reviewers,

Thank you again for taking the time to review our paper. Your questions and suggestions helped us clarify why STMix shares temporal components across event types and strengthen our comparisons with type-specific models.

We’ve replied to each review in its own thread and would welcome your thoughts on the revised paper.

Best regards,
The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.