acceptodds
Under review as a conference paper at ICLR 2027

From Rare Episodes to Complete Trajectories: Prior-Preserving Diffusion for Long-Tailed Time-Series Generation

Abstract

Long-tailed multivariate time-series generation is important for simulation, data sharing, and the development of reliable analysis models, yet infrequent tail states are often poorly represented in synthetic data. We identify a fundamental difficulty in window-based training, termed temporal rarity dilution: as the window length increases, the same finite tail episode appears in more highly overlapping windows without providing proportionally more independent tail information, while normal observations still dominate the learning objective. Existing time-series generators therefore struggle to preserve both sparse tail states and their temporal and multivariate context, particularly over long windows. To address this problem, we propose the Episode-to-Trajectory Diffusion Transformer (E2T-DiT), which separates when tail states occur from how their corresponding trajectories are realized. First, a duration-aware autoregressive planner models the occurrence, order, and persistence of lower-tail, normal, and upper-tail states through a compact state-duration representation. The sampled state segments are then expanded into time-aligned conditions for an Axial Diffusion Transformer, which generates complete trajectories by modeling temporal evolution and cross-variable interaction along their respective axes. Second, a prior-preserving training objective combines global denoising, state-balanced supervision, and boundary consistency, strengthening the realization of sparse tail states without altering their learned occurrence distribution. Experiments on four long-tailed time-series datasets with window lengths of 32, 64, and 128 show that E2T-DiT increases the average Correlational, Predictive, and Discriminative Scores by , , and , respectively, compared with the strongest baseline in each setting. It achieves the best in most comparisons, and preserves its advantage as the temporal context grows.

Primary area:
generative models
Keywords:
time series generation, diffusion model, long-tail distribution

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

This paper considers long-tailed time series generation and introduces E2T-DiT framework that separates discrete tail occurrence from continuous trajectory generation. A segment conditioned Axial Diffusion Transformer is used to generate complete multivariate trajectories conditioned on the sampled state sequence. It further introduces a state balanced denoising objective and penalty to strengthen learning on rare temporal states while preserving the empirical occurrence distribution learned by the separate state planner. While the paper addresses an interesting problem, I find that the current formulation, experimental design, and presentation do not provide sufficiently convincing evidence for its main claims.

Critical strengths
  1. The paper targets an important problem called rare temporal behavior generation in time series. The separation between state occurrence and trajectory realization is intuitively reasonable and could potentially be useful.
Critical weaknesses
  1. The experiment evaluation does not validate the core contribution. The paper is motivated by long tail generation and rare episodes, but the primary comparison in Table 1 relies on conventional metrics that are not specifically designed to measure rare state or rare episode fidelity. More importantly, the method explicitly claims to model several features including state occurrence, episode duration and lower/upper tail prevalence, yet it does not report baseline comparison for any of these quantities.

  2. The “temporal rarity dilution” is not convincingly demonstrated. The argument that the same rare episode appears in more overlapping windows as WW increases is intuitive, but no controlled experiment isolates this effect. While some baselines perform worse for larger windows and the paper interprets it as support for temporal rarity dilution, longer sequences are intrinsically more difficult to model due to increased temporal dependency length and higher dimensionality. Therefore, deterioration with larger WW does not uniquely support the proposed hypothesis.

Questions for authors
  1. What is the direct evidence that E2T-DiT generates rare episodes better than DLTTS and other baselines? Please provide more metrics on baseline comparisons rather than only global generation metrics.

  2. What exactly are c⋆c^\star, LL, and UU for SWaT? The paper defines SWaT tail points using attack labels, which does not seem compatible with Eq. (1) and the boundary consistency objective.

  3. What motivates the threshold of oil temperature <4<4 for ETTh1 and ETTm1? Is this based on domain knowledge or empirical quantile? How sensitive are the results to the threshold hyperparameter?

  4. Can the authors provide a more direct empirical validation of temporal rarity dilution? The performance degradation for longer windows could arise simply from difficult in long sequences modeling.

  5. Are the results averaged over multiple runs? Table 1 reports single values and I could not find variance or confidence intervals. Given the stochasticity of generative training and metrics, uncertainty results would help assess whether relatively small improvements are statistically reliable.

  6. Could the authors provide comparisons with recent methods such as [1][2][3]? The baseline comparison appears somewhat outdated and incomplete. The improvements over older baselines do not necessarily demonstrate competitiveness against the current methods.

[1] Population aware diffusion for time series generation.

[2] DiM-TS: Bridge the Gap Between Selective State Space Models and Time Series for Generative Modeling.

[3] TSGDiff: Rethinking Synthetic Time Series Generation from a Pure Graph Perspective.

Reviewer_dB3QOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies multivariate time-series generation with rare states, focusing on the occurrence, duration, and continuous trajectories of rare events. The proposed framework separates generation into two components: a state planner generates a sequence of states and their durations, and a conditional diffusion model generates multivariate trajectories conditioned on that sequence. State-balanced supervision and boundary-consistency constraints are introduced to improve the realization of rare states while preserving overall sequence quality. The main contribution is the combination of explicit event-duration modeling, conditional trajectory generation, and state-level training objectives. The reported evaluation uses multiple time-series datasets and generation-quality metrics.

Critical strengths

The paper targets a practically important challenge in time-series generation. Existing generative models tend to be dominated by frequent normal patterns, which can lead to poor modeling of rare but potentially important events. Focusing on rare-event generation addresses a meaningful limitation with clear relevance to applications such as industrial monitoring, anomaly analysis, and risk-sensitive forecasting.

Critical weaknesses
  1. The connection between the experimental datasets and the central motivation needs stronger substantiation. Defining minority states through thresholds on standard time-series datasets does not, by itself, establish the event-level learning difficulty motivating the paper. Aggregate generation improvements also do not establish that the proposed difficulty in long windows has been alleviated. Without a direct connection between these properties and the reported gains, the experiments may support the method’s utility without supporting its explanation of why the problem arises or why the method resolves it. This is the primary concern affecting my assessment.

  2. The proposed method combines a state planner, a duration representation, and a conditional trajectory generator with additional training objectives. It is unclear whether the reported benefits depend on these particular choices or whether comparable performance could be obtained by substituting other reasonable implementations, such as an alternative conditional generator.

Questions for authors
  1. Do the datasets exhibit the rare-event structure targeted by the method? Please clarify how state thresholds are defined and report tail prevalence, the number of independent events, and event-duration distributions in the training and test sets.

  2. Does author try difference window length? Please also specify the split boundaries and window strides, and clarify whether training and test windows share raw observations or contain different portions of the same rare event.

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

The paper proposes E2T-DiT for the generation of multivariate time series with rare temporal states. It separate generation into a Transformer that generates duration-aware autoregressive state prior and a segment-conditioned diffusion model that generates conditional measurements. The model also combines a state-balanced loss, which promotes learning of sparse states, and a soft boundary-consistency penalty, which helps generated values follow the sampled states. Experiments show the generation quality on four datasets and three window lengths.

Critical strengths
  1. The method separates episode occurrence and duration modeling from continuous trajectory generation using a Transformer by a conditional diffusion model. In this way, it can learn the discrete episode schedule separately and use it for continuous trajectory generation. This modeling choice is reasonable.
  2. The experiment section provides rich architecture and training detail and the results seems to be generally favorable. The ablation results provide useful information for experimental choices.
Critical weaknesses
  1. The method uses a duration-aware autoregressive planner to learn the empirical distribution of tail occurrence and duration. However, preserving this distribution does not necessarily mean that the generated trajectories faithfully follow the sampled plans. Erealize=∣rgen−rplan∣E_{\mathrm{realize}}=|r_{\mathrm{gen}}-r_{\mathrm{plan}}| (equation 14) only compares aggregate frequencies. The tail rate could match despite incorrect episode timing. Evaluating the alignment between the planned and generated states at each time step is therefore needed.
  2. It is true that Appendix A specifies the loss weights and conduct ablation study to compare including and removing the ρ\rho and λbnd\lambda_{bnd}. But it unclear how these weights are selected and how these values affect the generation quality. It would be helpful to include the sensitivity analysis on these two hyperparameters.
Questions for authors
  1. How accurately do generated trajectories follow sampled states at each time step with respect to episode timing and duration?
  2. For SWaT, how are attack labels defined and how are they mapped to the bound (L, U) states?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.