acceptodds
Under review as a conference paper at ICLR 2027

Context-Aware Diffusion in Temporal Event Streams

Abstract

In the context of continuous time event sequence prediction, diffusion models have achieved state-of-the-art performance. These methods propose to model the entire event sequence by simultaneously diffusing both the time intervals and the event types of the target sequence. However, they employ the default forward noising process, which does not take into consideration the local structural and temporal patterns observed in event sequences. In this work, we propose a locality-based structure-aware noising technique, which customizes the noise scale at each event, considering temporal patterns in the neighborhood. Building on this framework, we further develop specialized training and sampling procedures tailored to the proposed noising mechanism and provide theoretical analysis to support our design choices. We validate our approach in five real-world datasets while comparing several state-of-the-art baselines for both long- and short-horizon forecasting tasks. The proposed method achieves up to 20.11% in time prediction and up to 7.91% in type prediction over the most competitive baseline. Our quantitative and qualitative evaluations establish the effectiveness of data-driven locality-aware noising in diffusion models for continuous-time event sequences

Primary area:
generative models
Keywords:
temporal event streams, diffusion models, noise scaling

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_p4uwOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a context-aware diffusion framework for continuous time event sequence forecasting. The key idea is to introduce flexible noise scales that adapt to the temporal structure of individual events over the diffusion process. The experimental results show improvements on time predictions.

Critical strengths
  1. The key idea is clear and straight-forward.
  2. The improvements showed in experimental results are convincing.
Critical weaknesses
  1. The temporal heterogeneity of the experimental datasets is not characterized. Although some datasets are qualitatively described as sparse, it is difficult to determine whether larger performance gains actually correspond to greater temporal heterogeneity.
  2. Eq. (19) appears to contain a notation or formulation error on my side. May need a check.
Questions for authors
  1. How does ScaleDiff differ from existing data-dependent adaptive noising frameworks?
  2. Could the authors explain how the framework achieve “locality-aware”?
  3. Since ScaleDiff is specifically designed for datasets with significant variation in temporal event distributions, how does its performance change across datasets with different levels of temporal heterogeneity?
6 comments
AuthorsOpenReview ↗

Response to Reviewer p4uw

We thank the reviewer for their careful reading of our manuscript and for their constructive feedback. We appreciate the insightful comments and suggestions, which have helped us identify areas that require further clarification and improvement. We address each concern in detail below.We have uploaded a revised version of the manuscript, with updated and newly added text highlighted in blue for ease of reference.

W1 & Q3: The temporal heterogeneity of the experimental datasets is not characterized. Although some datasets are qualitatively described as sparse, it is difficult to determine whether larger performance gains actually correspond to greater temporal heterogeneity. Since ScaleDiff is specifically designed for datasets with significant variation in temporal event distributions, how does its performance change across datasets with different levels of temporal heterogeneity?

We thank the reviewer for asking the question of temporal heterogeneity in data.

We characterize inter-event timing using three global statistics: mean ($\mu$), standard deviation (σ\sigma), and coefficient of variation (CV=σ/μ\mathrm{CV}=\sigma/\mu). The mean quantifies the typical inter-event interval, the standard deviation measures absolute temporal dispersion, and CV measures dispersion relative to the mean. Statistics are computed over all inter event time values pooled within each dataset, including the initial event's gap; the global standard deviation uses the sample definition.

Lower RMSE Time is better. The relative improvement is

A−BA×100\frac{A-B}{A}\times100

where A=RMSETimeCDiffA=\mathrm{RMSE Time_{CDiff}}

B=RMSETimeScaleDiffB=\mathrm{RMSE Time_{ScaleDiff}}

Real-world datasetMean (μ\mu)Std (σ\sigma)Std/Mean (CV)CDiff RMSE TimeScaleDiff RMSE TimeRMSE reduction
Amazon0.52380.34130.6520.4640.34425.86%
Stack Overflow0.87531.08191.2361.1990.98318.02%
Taxi0.21850.29461.3480.3510.27521.65%
Retweet12.006816.90061.40827.10121.07422.24%
Taobao0.17340.43462.5060.5510.40526.50%

Performance across diverse real-world temporal distributions. We assess temporal heterogeneity using the global mean, standard deviation, and coefficient of variation (CV) of inter-event intervals. The five real-world datasets span markedly different temporal characteristics: global mean intervals range from 0.1734 (Taobao) to 12.0068 (Retweet), global standard deviations range from 0.2946 (Taxi) to 16.9006 (Retweet), and CV ranges from 0.652 (Amazon) to 2.506 (Taobao).

ScaleDiff consistently outperforms CDiff in temporal RMSE on all five real-world datasets, achieving reductions of 25.86% on Amazon, 18.02% on Stack Overflow, 21.65% on Taxi, 22.24% on Retweet, and 26.50% on Taobao. In particular, strong improvements are obtained at both relatively low temporal dispersion (Amazon: CV = 0.652) and high relative dispersion (Taobao: CV = 2.506). Retweet additionally demonstrates a substantial improvement despite having the largest absolute mean interval and standard deviation of the real-world group.

These results highlight ScaleDiff's consistent effectiveness across diverse real-world event-time distributions, encompassing distinct temporal scales and degrees of variability. The broad improvements support the practical value of ScaleDiff for temporal prediction under heterogeneous event-timing patterns. We provided the updated discussion in Appendix L.

W2: Eq. (19) appears to contain a notation or formulation error on my side. May need a check.

We thank the reviewer for identifying this error in the mathematical formulation. We acknowledge the oversight and clarify that the correct formulation is given by:

\operatorname{diag}\left( \frac{x_0-x_l}{x_u-x_l}(\kappa_u-\kappa_l)+\kappa_l \right).$$ We have corrected the corresponding equation in the revised manuscript to accurately reflect the intended formulation. We appreciate the reviewer's careful observation, which has helped us improve the mathematical clarity and correctness of the presentation.
AuthorsOpenReview ↗

Response to Reviewer p4uw (continued)

Q1: How does ScaleDiff differ from existing data-dependent adaptive noising frameworks?

We thank the reviewer for raising this question. The central novelty of ScaleDiff is not simply the introduction of adaptive noise, but the explicit connection between relative inter-arrival times, local temporal density, and event-specific corruption variance. Existing frameworks demonstrate that data-dependent or structured noising can improve diffusion. However, their adaptation mechanisms operate at different levels and are designed to capture different structural properties. In the following, we discuss the existing adaptive noising frameworks and demonstrate how ScaleDiff differs from them.

  1. MuLAN [1]: Input-dependent adaptation versus temporal-density-guided adaptation. MuLAN introduces multivariate, input-dependent diffusion schedules learned through a variational objective, enabling different noise rates across image coordinates. Thus, MuLAN is genuinely data-dependent. ScaleDiff differs in how this adaptation is determined. Rather than relying on a generative objective to learn an unrestricted coordinate-wise noise allocation, ScaleDiff explicitly constructs event-specific scale targets from relative inter-arrival times. This introduces a direct relationship between temporal proximity and corruption magnitude: short intervals in dense regions receive smaller scales, while longer intervals receive relatively larger scales. For temporal event streams, input dependence alone does not guarantee that noise allocation reflects local event density. MuLAN learns adaptive corruption from input characteristics; ScaleDiff explicitly supervises the adaptation using a temporal-density-based scaling rule. This provides a targeted inductive bias for protecting fine-grained event timing.

  2. Inverse Heat Dissipation [2]: Spatially structured corruption versus irregular temporal locality. Inverse Heat Dissipation employs heat-equation-based smoothing to progressively suppress high-frequency spatial information. Its forward operator exploits the regular geometry of image grids and is not dynamically adjusted according to the local density of individual input samples. In temporal point processes, consecutive event indices may correspond to milliseconds or hours of elapsed time. Therefore, applying an image-oriented smoothing operator over event positions would not automatically distinguish dense bursts from sparse regions. ScaleDiff instead determines the corruption magnitude from actual relative inter-arrival times, explicitly accounting for irregular physical-time separation rather than assuming uniformly spaced neighboring observations.

  3. Non-Isotropic Gaussian Diffusion [3]: General covariance flexibility versus data-dependent, locality-aware covariance. Non-Isotropic Gaussian Diffusion generalizes the forward process to Gaussian noise with a specified non-isotropic covariance. Its central contribution is mathematical flexibility in the noise distribution, rather than learning a separate covariance from each sequence's temporal-density structure. ScaleDiff goes beyond allowing non-uniform covariance by specifying what determines that covariance. Event-specific scales are constructed from relative inter-arrival times, ensuring that the target corruption variance explicitly reflects dense and sparse temporal regions. Thus, non-isotropic diffusion establishes that covariance can be structured, whereas ScaleDiff establishes a temporally meaningful rule for constructing sample-specific covariance. The distinction lies in explicitly incorporating event-time heterogeneity into the corruption mechanism.

  4. Blue Noise for Diffusion Models [4]: Spectrally structured corruption versus temporal-density-dependent variance. Blue Noise introduces time-varying, spatially correlated noise to exploit the spectral properties of images. Although its correlation structure changes across diffusion timesteps, it is not designed to estimate separate noise magnitudes from the dense and sparse regions of each input event sequence. ScaleDiff instead adapts the marginal Gaussian variance associated with individual inter-arrival times. This distinction is particularly important because neighboring event indices do not imply fixed physical-time separation. While Blue Noise addresses how corruption is correlated across coordinates, ScaleDiff addresses how strongly each event interval should be corrupted according to its local temporal density.

[1] Sahoo et al 2024 Diffusion Models With Learned Adaptive Noise NeurIPS 2024

[2] Rissanen et al 2023 Generative Modelling With Inverse Heat Dissipation ICLR 2023

[3] Voleti et al 2022 Score-based Denoising Diffusion with Non-Isotropic Gaussian Noise Models NeurIPS Workshop 2022

[4] Huang et al 2024 Blue Noise for Diffusion Models ACM SIGGRAPH 2024

AuthorsOpenReview ↗

Response to Reviewer p4uw (continued)

Continuing Q1 response here.

What makes ScaleDiff methodologically distinct?

ScaleDiff introduces a complete locality-aware diffusion framework that explicitly connects temporal structure to event-specific noise variance.

First, relative inter-arrival times determine the initial scale targets, establishing a direct relationship between dense and sparse regions and their intended corruption magnitudes.

Second, these scales are incorporated into the forward Gaussian covariance.

Unlike a global scalar schedule, the diagonal matrix (StS_t) allows different intervals within the same sequence to receive different noise variances at the same diffusion timestep.

Third, a separately trained scale prediction module learns to recover the initial scale targets from historical context, noisy intermediate states, and timesteps, enabling the adaptation during inference without access to future ground-truth intervals.

Finally, the predicted scales are incorporated into the corresponding Gaussian reverse transitions, allowing the proposed event-specific scaling mechanism to operate throughout sampling.

The distinction from existing data-dependent approaches is therefore precise - learning input-dependent corruption. ScaleDiff uniquely targets relative inter-arrival-time heterogeneity through sample-specific, event-wise variance scaling.

This design directly addresses the unequal relative impact of Gaussian perturbations on short and long inter-arrival times. Our region-wise reconstruction analysis, synthetic experiments, and forecasting evaluations across five real-world datasets provide empirical support for this locality-aware formulation. In summary, data-dependent noising alone does not imply locality-aware noising. MuLAN adapts according to learned input characteristics, and the other vision-based methods exploit prescribed structural properties. ScaleDiff explicitly identifies local temporal density as the property governing corruption and incorporates it into the forward covariance and scale-aware reverse process. This is the central methodological distinction of our work. We updated the Related Works in summary in Appendix D

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

The paper proposes ScaleDiff, aiming to improve long-term domain prediction for continuous-time event streams. Existing diffusion models typically use a unified Gaussian noise, but the paper argues that event intervals have dense and sparse regions, and uniform noise may prematurely destroy the temporal structure in dense regions. To address this, the authors construct per-event noise scales S0S_0 based on the target event intervals, then use a scale prediction network to estimate these scales during inference, and modify the forward diffusion process and the reverse denoising process accordingly.

Critical strengths
  1. Irregular time event streams indeed exhibit significant differences in temporal density. Using different noise scales for dense and sparse events is a reasonable and research-worthy direction.

2.The experiments cover a broad range, and the paper not only compares multiple existing baselines but also conducts various validation experiments.

Critical weaknesses
  1. The training and inference process of the scale prediction network are inconsistent. If Algorithm 1 is the actual training process, the model mainly learns the scale from clean target sequences, while noisy sequences need to be processed during inference. This introduces a distribution shift between training and inference. If the actual training uses xtx_t, then the algorithm pseudocode does not accurately describe the real implementation.

  2. There are some obvious AI traces, which are not convenient for reading. In particular, the images are clearly AI-generated, but this is not mentioned in the AI declaration.

Questions for authors
  1. In Algorithm 1, the input is (xc,x0,0)(x_c,x_0,0), while equations (18), Appendix G, and Algorithm 2 use (xc,xt,t)(x_c,x_t,t). Which training protocol is actually used in the experiments? How to ensure that the reverse sampling still conforms to the derived posterior form when re-estimating S0S_0 at each step?

  2. Can you explain the proof of Property 1 and strictly distinguish between the terminal condition distribution and the marginal distribution? For the scale prediction error, can you provide an argument on error propagation or distribution convergence throughout the entire backward chain?

5 comments
AuthorsOpenReview ↗

Response to Reviewer dDUP

We thank the reviewer for their careful reading of our manuscript and for their constructive feedback. We appreciate the insightful comments and suggestions, which have helped us identify areas that require further clarification and improvement. We address each concern in detail below. We have uploaded a revised version of the manuscript, with updated and newly added text highlighted in blue for ease of reference.

W1: The training and inference process of the scale prediction network are inconsistent. If Algorithm 1 is the actual training process, the model mainly learns the scale from clean target sequences, while noisy sequences need to be processed during inference. This introduces a distribution shift between training and inference. If the actual training uses x_t, then the algorithm pseudocode does not accurately describe the real implementation.

We respectfully disagree with the reviewer's characterization of our framework as suffering from a train–test distribution mismatch. We believe this concern arises from a misunderstanding of the two distinct training stages: (i) independent training of the scale prediction module and (ii) subsequent training of the diffusion model using the pretrained scale predictor.

First, the scale prediction module fθsSf_{\theta_s}^S is trained independently and is frozen throughout Algorithm 1 and 2. As described in Eq. (18), the scale module is trained using randomly sampled diffusion timesteps, receiving �0��0� as input, including the clean-state configuration �1��1�. Therefore, the module is not trained exclusively on clean targets; it explicitly learns to estimate scales from both clean and noisy states. During Algorithm 1, the frozen pretrained fθsSf_{\theta_s}^S is used only for scale prediction. Thus, Algorithm 1 correctly represents the training of the diffusion backbone, not the training procedure of the scale predictor itself.

Second, we directly evaluate the reviewer's concern through an additional controlled ablation. Specifically, we compare our original configuration, which uses fθsS(xc,x0,0)f_{\theta_s}^S (x_c,x_0,0) (denoted by ScaleDiff x0 here) during diffusion training, with an alternative that uses fθsS(xc,xt,t)f_{\theta_s}^S (x_c,x_t,t) (denoted by ScaleDiff xt here), matching the input configuration employed during inference. Both configurations use the same independently pretrained scale prediction module.

DatasetMethodOTDRMSE TypeRMSE TimesMAPE
TaoBaoScaleDiff xt42.150± 0.3112.452 ± 0.0790.405±0.001139.144±0.805
TaoBaoScaleDiff x042.081±0.2842.443±0.0650.405±0.002139.708±1.712
RetweetScaleDiff xt58.808 ± 0.2312.917 ± 0.05621.107 ± 0.07185.2868 ± 0.264
RetweetScaleDiff x058.832 ± 0.2782.901 ± 0.06021.074 ± 0.05485.252 ± 0.349
TaxiScaleDiff xt19.121 ± 0.6701.052 ± 0.0670.277 ± 0.00477.123 ± 1.232
TaxiScaleDiff x018.767 ± 0.3741.013 ± 0.0560.275 ± 0.00376.631 ± 0.814
AmazonScaleDiff xt37.298 ± 0.2382.143 ± 0.0240.344±0.00277.920±0.560
AmazonScaleDiff x037.345 ± 0.2282.155 ± 0.0200.344 ± 0.00278.203 ± 0.672
StackOverflowScaleDiff xt38.805 ± 0.2701.532 ± 0.0230.979 ± 0.00686.339 ± 0.225
StackOverflowScaleDiff x038.918 ± 0.5561.501 ± 0.0570.983 ± 0.01286.492 ± 0.185

The results across five datasets demonstrate largely comparable performance between the two configurations. The absolute differences in RMSE-Time are at most 0.004, while sMAPE differences remain below 0.5. Neither configuration consistently dominates: for OTD, the noisy-state configuration performs better on three datasets, whereas the clean-state configuration performs better on two. Event-type prediction also remains broadly comparable across datasets. These observations indicate that matching the scale-prediction inputs between diffusion training and inference does not consistently improve forecasting performance. We updated the same in Appendix M.

AuthorsOpenReview ↗

Response to Reviewer dDUP (continued)

W2: There are some obvious AI traces, which are not convenient for reading. In particular, the images are clearly AI-generated, but this is not mentioned in the AI declaration.

We thank the reviewer for raising this concern. However, we would like to clarify that the figures presenting our experimental results were created by the authors using our own data, without the use of AI-based image-generation tools. For some figures, we consulted ChatGPT to obtain general Python plotting code templates, which we subsequently adapted and executed using our own experimental values. ChatGPT was not used to generate, modify, or manipulate any experimental results or their visual representations. Regarding the main illustration, ChatGPT was used to assist in preparing the figure based on the content of our paper and a diagram previously presented by the authors. We acknowledge that this assistance should have been made clearer in the AI Use Statement. We have therefore revised the AI-use statement to explicitly describe the limited role of ChatGPT in figure preparation and Python code assistance. We hope this clarification addresses the reviewer's concern.

AI Use Statement

The authors used the freely available versions of ChatGPT and Claude solely to assist with manuscript preparation, including restructuring, language refinement, grammatical corrections, and improving clarity and readability. AI tools were also used to assist in preparing the main conceptual illustration based on the authors' original research ideas and a diagram developed by the authors.

Additionally, AI tools were used to generate Python plotting code for experimental visualizations. The authors provided their own experimental data, incorporated the corresponding values into the generated code, and executed it to produce the experimental plots. Thus, while AI assisted with the plotting code, all experimental data, results, and visualizations were based entirely on the authors' original experiments. No AI-based image-generation tools were used to generate experimental plots.

No generative AI tools were used to develop the research methodology, design or conduct experiments, generate or modify experimental results, perform ablation studies, select citations, or formulate scientific conclusions. All AI-assisted content was carefully reviewed and verified by the authors, who take full responsibility for the methodology, experiments, results, interpretations, citations, and overall scientific integrity of the manuscript.

AuthorsOpenReview ↗

Response to Reviewer dDUP (continued)

Q1: In Algorithm 1, the input is (xc,x0,0), while equations (18), Appendix G, and Algorithm 2 use (xc,xt,t). Which training protocol is actually used in the experiments?

We use (xc,x0,0)(x_c,x_0,0) for training and the addressal of train-test mismatch is mentioned in response to W1.

Q1 (contd.): How to ensure that the reverse sampling still conforms to the derived posterior form when re-estimating at each step?

We would like to emphasize that step-wise scale re-estimation preserves the analytically derived Gaussian transition parameterization. At each timestep, the independently trained scale prediction module estimates the initial scale matrix (S_0) from the current noisy state and historical context. Both (StS_t) and (St−1S_{t-1}) are then consistently computed from this estimate using Eq. (10), followed by the corresponding covariance and reverse mean through Eqs. (11), (15), and (17). Thus, the reverse transition retains its derived Gaussian form throughout sampling. Importantly, the scale prediction network is explicitly trained across randomly sampled diffusion timesteps to estimate the same underlying initial scale matrix (S0S_0). Consequently, re-estimation at each step refines the estimate of a common scaling quantity using the available noisy state, rather than introducing unrelated scaling factors.

We perform an experiment where we find the cumulative percentage recovery on Stackoverflow dataset for CDiff and ScaleDiff for each denoising step shown in Figure 11 of Appendix S. Our empirical analysis further demonstrates the effectiveness and stability of this mechanism. Across the reverse diffusion trajectory, ScaleDiff achieves 7.33% cumulative error reduction, substantially exceeding the 0.36% observed for CDiff. Furthermore, ScaleDiff's relative RMSE advantage increases from 6.21% to 12.77% as denoising progresses. These results demonstrate that step-wise scale re-estimation is accompanied by sustained error reduction and progressively improved alignment with ground-truth temporal sequences.

Taken together, the analytical formulation and empirical evidence establish that ScaleDiff maintains the derived Gaussian transition structure at each reverse step while achieving substantially more effective progressive refinement than conventional diffusion. The results reinforce the effectiveness of our adaptive scale-estimation strategy and its contribution to improved reverse diffusion dynamics.

Reviewer_KKaZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper targets the problem of continuous-time event sequence prediction. The authors introduce a locality-based, structure-aware noising technique, namely ScaleDiff, which customizes the noise scale at each event, considering temporal patterns in the neighborhood. Extensive experiments on five real-world datasets demonstrate the effectiveness of the proposed solution.

Overall, the paper targets an important problem, and the motivation is clear. The related work section is generally well explained. The authors provide experiments on five real-world datasets. However, the technical strength of the paper is relatively unclear since most of the design of the proposed method seems similar to the existing method (e.g., CDiff).

Additional comment: The font size of Figure 1 is too small. I think it would be better to enlarge this figure.

Critical strengths

(S) The paper targets an important problem

(S) The authors conduct experiments on five real-world datasets

Critical weaknesses

(W) The technical contribution of this paper is somewhat unclear

(W) The differences and novelty compared to other papers should be more clearly explained

Questions for authors

Minor question: What is the difference between “continuous-time event sequences” and “event streams”? I think the term “stream” sounds like online processing or incremental learning. I would like to know why the authors use the term “stream” in this context.

5 comments
AuthorsOpenReview ↗

Response to Reviewer KKaZ

We thank the reviewer for their careful reading of our manuscript and for their constructive feedback. We appreciate the insightful comments and suggestions, which have helped us identify areas that require further clarification and improvement. We address each concern in detail below. We have uploaded a revised version of the manuscript, with updated and newly added text highlighted in blue for ease of reference.

W1 : The technical contribution of this paper is somewhat unclear

We thank the reviewer for pointing this out. Here we briefly clarify our technical contribution. Unlike image data, temporal point processes exhibit substantial variations in event density and inter-event time distributions, even within the same sequence. In particular, sparse and dense temporal regions can have significantly different variance characteristics. Directly borrowing a conventional noising scheme from the image domain does not explicitly account for these variations. Consider three temporal intervals within a sequence, characterized by different event frequencies. Dense regions, containing frequent events, can lose their temporal patterns rapidly under the conventional diffusion process, whereas sparse regions may retain their patterns for longer. This creates an imbalance in the forward diffusion process: high-variance regions remain informative for more diffusion steps, while low-variance regions rapidly become dominated by noise. As a result, the reverse model may learn to reconstruct high-variance regions more effectively while struggling with dense, low-variance regions, potentially leading to suboptimal estimation of the noise prediction network (ϵθ\epsilon_\theta) (Eq. 6).

We therefore propose context-adaptive variance scaling at early diffusion steps to regulate the rate at which temporal information is corrupted across regions with different statistical characteristics. Our objective is to preserve the relative temporal structures and patterns more consistently during diffusion, thereby facilitating the learning of reverse diffusion dynamics.

To make this motivation clearer, we have added an illustrative Figure 6 in Appendix F comparing the distortion of temporal patterns under conventional noising and our proposed variance-adaptive noising. To investigate this, we conduct an experiment on the Retweet dataset to evaluate how well ScaleDiff and CDiff preserve event-index correspondence across consecutive forward diffusion steps. Specifically, for each target sequence of length 20, we consider the absolute timestamps of events at diffusion steps (T) and (T+1). For each event at step (T+1), we identify its nearest event at step (T) based on the minimum absolute timestamp difference. We then check whether the nearest event retains the same index (i) across both diffusion steps. For each sequence, we compute the percentage of events whose nearest matches retain the same index. We then average these percentages across all sequences in the dataset for each pair of consecutive diffusion steps. The resulting average percentage is plotted against the progression of forward diffusion steps. As illustrated in the figure, ScaleDiff consistently achieves a higher average percentage of index matches than CDiff throughout the diffusion process, with a particularly pronounced difference in the early and intermediate steps. This suggests that ScaleDiff better preserves event-index correspondence and the relative temporal structure of event sequences under increasing noise, whereas CDiff exhibits a more rapid loss of correspondence. The illustration demonstrates how conventional noising can disrupt temporal structures at different rates, whereas the proposed mechanism provides more balanced corruption across intervals.

Importantly, this modification is not merely a change to the forward noise schedule. It also requires a corresponding modification to the reverse diffusion. We therefore introduce a scale-prediction network that estimates the required scaling factors and incorporates them into reverse-process reconstruction.

These modifications collectively constitute the technical contribution of our approach: a context-adaptive diffusion mechanism designed to account for the heterogeneous variance characteristics of temporal point processes.

AuthorsOpenReview ↗

Response to Reviewer KKaZ (continued)

W2 : The differences and novelty compared to other papers should be more clearly explained

We thank the reviewer for raising the issue.

The novelty of ScaleDiff does not lie merely in employing non-uniform Gaussian noise, but in explicitly deriving event-specific diffusion scales from relative inter-arrival times and incorporating these scales into both forward corruption and reverse sampling. Here we briefly discuss the novelty of our method with respect to existing works.

Among vision-based models, MuLAN [1] learns multivariate, input-dependent noise schedules for image generation, whereas ScaleDiff introduces inter-arrival-time-dependent inductive bias to regulate corruption in irregular temporal sequences.

Inverse Heat Dissipation [2] progressively removes image details through heat-equation-based spatial smoothing, exploiting the spatial organization of images. However TPPs are irregularly sampled sequences without a fixed temporal grid. Consecutive events in event-index space may be separated by milliseconds or several hours. ScaleDiff explicitly uses actual inter-arrival times to determine the corruption magnitude.

Non-Isotropic Gaussian Diffusion [3] generalizes conventional diffusion by allowing Gaussian perturbations with non-isotropic covariance structures. However, the generalization to non-isotropic covariance does not, by itself, specify how the covariance should reflect the temporal structure of an irregular event sequence.

Blue Noise for Diffusion Models [4] introduces temporally varying correlated noise to exploit the spectral characteristics of image signals. Its primary focus is on modifying the correlation structure of noise across spatial locations and diffusion steps. In contrast, ScaleDiff adapts the marginal noise variance associated with individual temporal intervals according to their relative magnitudes.

Among diffusion-based temporal point process methods, ADD-AND-THIN [5] adopts a fundamentally different forward diffusion strategy where it corrupt the input sequence through event thinning and homogeneous Poisson superposition. In contrast, ScaleDiff retains the event sequence structure and introduces Gaussian perturbations to inter-arrival times, with the perturbation variance determined by relative temporal intervals.

Specifically, ScaleDiff introduces the following components:

  1. Inter-arrival-time-dependent scale construction: Event-specific scale targets are derived from relative inter-arrival times, explicitly distinguishing densely and sparsely occurring events.

  2. Locality-aware Gaussian covariance: The derived scales determine the variance of Gaussian perturbations in the forward process. This reduces the relative corruption of short inter-arrival times while allowing larger perturbations for longer intervals.

  3. Learned scale estimation for inference: A dedicated scale-prediction network estimates the required initial scaling factors from historical context and noisy intermediate states. This enables adaptive reverse sampling without requiring access to future ground-truth intervals.

  4. Scale-aware reverse transitions: The predicted scales are incorporated into the reverse diffusion transitions, ensuring consistency between the proposed forward corruption mechanism and the corresponding reverse reconstruction process.

In summary, ScaleDiff is not simply an application of adaptive image diffusion to TPPs, nor is it a variation of event-thinning-based TPP diffusion. Its principal contribution is the explicit mathematical coupling of relative inter-arrival times, local temporal density, and event-specific Gaussian corruption variance, together with the scale-aware reverse mechanism required to realize this formulation during sampling. This provides a temporal-structure-driven approach to diffusion in irregular event sequences. We updated the Related Works in Appendix D

[1] Sahoo et al 2024 Diffusion Models With Learned Adaptive Noise NeurIPS 2024

[2] Rissanen et al 2023 Generative Modelling With Inverse Heat Dissipation ICLR 2023

[3] Voleti et al 2022 Score-based Denoising Diffusion with Non-Isotropic Gaussian Noise Models NeurIPS Workshop 2022

[4] Huang et al 2024 Blue Noise for Diffusion Models ACM SIGGRAPH 2024

[5] Ludke et al 2023 Add and Thin: Diffusion for Temporal Point Processes NeurIPS 2023

AuthorsOpenReview ↗

Response to Reviewer KKaZ (continued)

Q1: Minor question: What is the difference between “continuous-time event sequences” and “event streams”? I think the term “stream” sounds like online processing or incremental learning. I would like to know why the authors use the term “stream” in this context

We agree that the term event stream is sometimes used in the literature to refer to events processed sequentially in an online or incremental learning setting. In this paper, however, we use event stream to refer to a temporally ordered sequence of events occurring at continuous-valued timestamps. Specifically, it does not imply online processing, incremental learning, or real-time event ingestion.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Discussion Phase

Dear AC, SAC, and PC

We thank the reviewers for their thoughtful evaluation. We addressed the feedback through clearer explanations, mathematical revisions, and new experiments. We highlight the key details for individual issues and responses to reviewers. Changes are marked in blue in the updated manuscript.

1. Summary of Reviewer Concerns Addressed (Reviewer ID, Question)

  • Technical contribution (KKaZ, W1): Clarified ScaleDiff's core motivation of adapting diffusion noise to inter-event times rather than treating dense and sparse temporal regions uniformly. Figure 6 (Appendix F) demonstrates how our adaptive noising better preserves temporal patterns, supporting its use for event forecasting.

  • Novelty and data-dependent noising (KKaZ, W2) (p4uw,Q1) : Clarified ScaleDiff's distinction from adaptive noising methods in vision and homogeneous Poisson-based temporal diffusion. ScaleDiff uses a learnable, context-aware scale module conditioned on historical context, the noisy/clean target sequence, and the diffusion timestep to estimate inter-event time-dependent scaling factors. Updated the related work section accordingly.

  • Scale module and train–test consistency (dDUP,W1 & Q1): Clarified the scale module's training and its use during diffusion backbone training. Additional experiments comparing inputs (xc,x0,0)(x_c,x_0,0) and (xc,xt,t)(x_c,x_t,t) showed no consistent performance advantage for either configuration, suggesting no meaningful discrepancy. Also clarified that the module is trained across noisy target sequences leading to no train-test mismatch. (Appendix M)

  • AI Use Statement (dDUP,W2): Clarified that figures were generated using Python scripts, with AI assistance for portions of plotting code. Acknowledged AI assistance in creating the main figure using the draft manuscript and an initial draft figure. Updated AI Use Statement. No AI image-generation tools were used, and AI was not used to fabricate values, select results, or infer plots from data.

  • Event stream definition (KKaZ, Q1): Clarified that an event stream is a time-ordered sequence of events with continuous-valued timestamps, without necessarily implying online or incremental processing.

  • Denoising-step analysis (dDUP, Q1): Added inter-event-time RMSE analysis across denoising steps, showing improvement for ScaleDiff and the opposite trend for CDiff (Appendix S).

  • Property 1 and scale-prediction error (dDUP, Q2): Provided the explanation of Property 1 and the scale-prediction error analysis already (Appendices A and B).

  • Temporal heterogeneity (p4uw, W1 & Q3): Analyzed temporal heterogeneity using the mean, standard deviation, and coefficient of variation of inter-event times, supporting ScaleDiff's effectiveness across heterogeneous sequences (Appendix L).

  • Mathematical formulation (p4uw, W2): Corrected the mathematical formulation following the reviewer's suggestion.

  • Achieving ‘Locality aware’(p4uw, Q2): Added Figure 5 (Appendix F) comparing dense-region preservation in ScaleDiff and CDiff, and clarified coarse (sparse) versus refined (dense) temporal features and their differences from vision and language.

2. Reviewer follow-up and request to the Area Chair

Following our rebuttal, reviewer p4uw acknowledged our detailed response and additional experiments, and reviewer KKaZ acknowledged our explanations. Both maintained their original scores, without specifying further technical questions or identifying particular unresolved concerns in their follow-up comments. At the time of writing, reviewer dDUP has not posted a follow-up response.

We respect the reviewers' assessments and their overall judgment. However, our added experiments directly examine temporal preservation during noising, better denoising behavior, performance invariant of temporal heterogeneity, and scale-module input consistency.

We therefore respectfully ask the Area Chair (AC) to consider the substantive rebuttal evidence and manuscript revisions alongside the original reviews, particularly where reviewers retained their scores without outlining which technical reservations, if any, remained unresolved after our responses. We request this new evidence be taken into consideration before the final assessment.

3. Conclusion

We believe the revision provides a clear depiction of why context-aware diffusion is needed, how ScaleDiff differs from previous method, and what added experiments show about its behavior. We sincerely appreciate your time and consideration and hope these technical clarifications and empirical findings will be given appropriate weight in the final decision.

Sincerely,
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.