acceptodds
Withdrawn from ICLR 2027

End-to-end Early Classification of Time Series in Non-Stationary Environments

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Early Classification of Time Series (ECTS) requires making accurate decisions as early as possible in inherently online and evolving environments. Yet, most existing methods assume stationarity and rely on separable designs, where classification and triggering are optimized independently, an assumption that fundamentally limits their adaptability under drift. In this work, we challenge this paradigm and study ECTS under non-stationary conditions. We provide the first controlled empirical comparison between separable and end-to-end approaches under multiple drift regimes. Building on Reinforcement Learning, we introduce DQeND, a unified architecture that jointly learns representation, classification, and triggering decisions while remaining directly comparable to state-of-the-art separable baselines. Across a wide range of drifts, DQeND demonstrates strong robustness across various non-stationary scenarios. Overall, our results indicate that end-to-end learning can offer improved adaptation capabilities for ECTS in dynamic environments and motivate further investigation of alternatives to separable designs.

Primary area:
learning on time series and dynamical systems
Keywords:
time series, reinforcement learning, online learning

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

The paper studies Early Classification of Time Series (ECTS) when the deployment distribution drifts, and asks whether jointly optimising classification and triggering beats the dominant separable paradigm (encode → classify → trigger), where the two components are trained independently. It proposes DQeND, an architecture that couples a frozen random-convolution ELM/ROCKET-style encoder with a Deep Q-Network whose action space is extended to {wait, predict class 1, …, predict class K}, so that a single RL agent decides both when to stop and which class to emit; training is offline and adaptation is online (prequential), with a re-introduced ε-greedy exploration. Evaluation uses a controlled drift simulator (binarised MNIST-1D) with four drift regimes — incremental location, incremental noise, abrupt location, and recurrent abrupt meta-class shuffling — plus one real corpus (arXiv abstracts) reduced to a binary majority-vs-rest task, with metrics decomposed into cumulative cost, hold-out cost on drifted data, and hold-out cost on the original distribution.

Critical strengths
  • The prequential protocol with hold-out cost on drifted versus original data is a clean, reusable way to separate adaptation from retention.
  • DQeND is deliberately architecture-matched to Alert (same DQN head, same frozen ELM encoder), so the comparison is not confounded by model size.
  • The reported findings are informative beyond the model: frozen triggers are catastrophic under drift, independently updating the two modules can be worse than not adapting, and RL trigger policies collapse under repeated abrupt shuffling.
Critical weaknesses
  1. DQeND changes three things at once (shared frozen encoder, joint action space, full gradient flow) and the only ablation cannot separate them, so the end-to-end paradigm itself is not shown to be the cause.
  2. The evidence is small and self-generated (binary MNIST-1D simulator, 1,000 training series, 5 seeds, no confidence intervals or tests) plus one arXiv task collapsed to majority-vs-rest, with no standard ECTS benchmark.
  3. The headline DQeND-versus-ELECTS/EARLIEST gaps are a few tenths of a cost unit with no significance testing, while α = 0.8 and the exploration rate ε = 5% are fixed and never ablated.
  4. Under recurrent abrupt meta-class shuffling ELECTS wins and DQeND's trigger policy collapses toward early predictions, a failure mode that is neither diagnosed nor mitigated.
  5. The RL formulation is asserted rather than derived (no MDP, no argument that the shaped reward equals the risk in Eq. 3/4), the drift schedule is described inconsistently, and efficiency is claimed with no cost or latency measurement.
Questions for authors
  1. DQeND changes three things at once (shared frozen encoder, joint action space, full gradient flow) and the only ablation cannot separate them, so the end-to-end paradigm itself is not shown to be the cause.
  2. The evidence is small and self-generated (binary MNIST-1D simulator, 1,000 training series, 5 seeds, no confidence intervals or tests) plus one arXiv task collapsed to majority-vs-rest, with no standard ECTS benchmark.
  3. The headline DQeND-versus-ELECTS/EARLIEST gaps are a few tenths of a cost unit with no significance testing, while α = 0.8 and the exploration rate ε = 5% are fixed and never ablated.
  4. Under recurrent abrupt meta-class shuffling ELECTS wins and DQeND's trigger policy collapses toward early predictions, a failure mode that is neither diagnosed nor mitigated.
  5. The RL formulation is asserted rather than derived (no MDP, no argument that the shaped reward equals the risk in Eq. 3/4), the drift schedule is described inconsistently, and efficiency is claimed with no cost or latency measurement.
Reviewer_dSJSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper compares end-to-end and separable paradigms for early time series classification (ECTS) under non-stationary conditions. Existing end-to-end methods are rarely evaluated against strong separable baselines, while non-stationarity research mainly focuses on separable methods with fixed classifiers. The paper proposes DQeND, combining a randomly initialized 1D convolutional encoder with a Deep Q-Network (DQN) to decide when to predict and support online adaptation. Experiments on synthetic distribution shifts and a text-derived time series dataset show competitive performance in several settings, but results vary across shift types and computational costs are relatively high.

Critical strengths

S1. The paper investigates early time series classification under non-stationary conditions, where distribution shifts can affect both classification accuracy and prediction timing. S2. DQeND integrates feature extraction and sequential prediction decisions into a single reinforcement learning framework. This allows the model to adapt its representations and stopping decisions together, potentially capturing interactions that are difficult to account for when these components are designed separately.

Critical weaknesses

W1. DQeND combines a randomly initialized 1D convolutional encoder with a DQN-based decision module. The paper does not sufficiently clarify the technical advances beyond this combination, nor does it explain why the proposed design is preferable to existing reinforcement learning-based ECTS approaches. W2. The methods compared appear to differ in the feedback available during online adaptation. If some baselines can use the full labeled time series, while reinforcement learning-based methods can only observe the prefix before prediction, such a comparison may favor certain methods.

Questions for authors

1.The authors should clarify the feedback and labeled data available to each baseline during online adaptation. Would the main conclusions hold if all methods were evaluated under the same feedback protocol? 2.What are the key technical advantages of DQeND over existing reinforcement learning-based ECTS methods? 3.In Table 6, DQeND incurs a higher cumulative cost than ELECTS under the A_class scenario. What causes this performance gap, and under what conditions does DQeND provide consistent advantages over existing methods?

Reviewer_ZSG4OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a new method for early classification of time series. The method acts under distribution shift settings, and is based on a convolutional feature extractor and an RL-inspired decision head that is trained end-to-end. The method is evaluated on synthetic datasets and on a real-world dataset (ArXiv abstracts), and compared to separable baselines.

Critical strengths
  • The distribution shift setting is interesting and relevant for real-world application, and is under tackled in the ECTS literature
  • The paper is well written and the description of the method is sound
Critical weaknesses
  • The novelty of the work is not clearly stated, and the paper does not make a clear case for why the proposed method is better than existing methods (ELECTS, Renault et al., 2026b)
    • Authors decided to delay related works to the appendix, which is an awkward choice given the limited information available in the intro about related works
      • First, the intro is not crystal clear about the fact that E2E learning for early classification of time series already exists
      • Typically, ELECTS seems like a relatively close work, but with a different architecture and loss, and it would be interesting to compare the two approaches in more detail, eg. using a common feature extraction architecture so that the comparison relies solely on the loss formulation / head architecture. This is done for the comparison to Alert, but not for ELECTS
      • Similarly, the intro does not tell much about how (Renault et al., 2026b) deals with non stationarity, what findings it has, and how it compares to the approach in this work
    • Looking at Fig1, and knowing that in the ELM encoder, the Conv1D is made of (freezed) random convolutions, this layout is very close to the separable one, in the sense that if one would move the linear layer from the left block to the right, then we would have a feature extractor made of handcrafted features (convolutions) and a decision head that is learned on top of it: are we just looking at yet another feature extractor for separable architectures (but with an RL head)? This would be a valid approach, but then the novelty of the work is not the end-to-end learned pipeline, but rather plugging a ROCKET feature extractor in (Renault et al., 2026a), up to an extra linear layer
  • Evaluation is too limited to bring strong conclusions
    • Evaluation is only performed in controlled settings (even the ArXiv task is very synthetic in the way it is built), those settings are interesting but cannot be sufficient, and it would be a good idea to have an intuition on the impact of the different parts of the methods
      • The feature extractor could be evaluated on standard benchmarks, and the same applies for the RL-inspired head
      • Only the ϵ\epsilon-greedy adaptation strategy is specific to distribution shift
    • Isn't it a bit of a problem to have deployment scenarios that do (a) "Pattern shifts smoothly from the beginning to the end of the series" and (b) "Pattern moves from the first to the last quarter after five steps" given that your method only does convolution + global mean pooling, hence location does not matter (except at the boundaries)? This seems like a super simple task for your architecture, and those are 2 scenarios among 4 for classification, and for triggering, it just shifts the decision time but not the pattern to be detected
    • Also, the ArXiv task consists in identifying if an abstract is from the majority class, but the majority class changes abruptly at a given point in time, such that the classifier trained on previous data cannot be used reliably to predict the label of the new data. This is a very adversarial setting, and it's unclear how it could relate to a real deployment scenario
  • Minor remarks:
    • In Eq3, it is a bit of a pity that the dependency of t^\hat{t} on ss and xx is not explicitly shown
    • p.4, l.179: isn't it Conv1D$_i$?
    • after Eq 6: "where −∆Cd(t) = Cd(t)−Cd(t−1)": isn't it Cd(t)−Cd(t−1) = ∆Cd(t)? (without the minus sign)
Questions for authors
  • Can you clarify the novelty of your work?
  • Can you evaluate which parts of your method are key for the success of your work?

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

Related papers

Discussion (0)

Sign in to comment.