acceptodds
Under review as a conference paper at ICLR 2027

FlexTempRule: Interpretable Time Series Classification via Learning Flexible Temporal Logic Rules

Abstract

Interpretable time series classification is a significant task for safety-critical systems, as it can help experts gain insight of the system behaviors, especially when the model is intrinsically interpretable and provides intuitive yet rigorous rationales that reflect its actual execution. However, existing rule-based approaches either fail to guarantee model–rationale consistency or restrict the evaluation scope of each learned logical rule to a predefined fixed interval. These limitations inevitably hinder the capture of flexible temporal correlation properties that may be satisfied at different regions across time series. In this paper, we propose a novel neuro-symbolic model called FlexTempRule, which supports adaptive learning of both fixed and slidable temporal logic rules to achieve interpretable time series classification with model-consistent rationales. Specifically, the time series is first dynamically divided into patches of varying lengths, expecting that the boundary points do not break temporal continuity. Then, three types of temporal logic rules with different degrees of sliding flexibility, as well as statistical features, are automatically learned and adaptively weighted to form advanced features for classification. Comprehensive experiments on real-world univariate and multivariate datasets in safety-critical domains show that FlexTempRule attains superior accuracy among rule-based baselines and competitive performance against other interpretable methods. More importantly, discriminative temporal logic rules are explicitly provided, along with their flexibility nature, thereby enhancing user trust in the model and deepening understanding of specific scenarios.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
interpretable time series classification, flexible rule learning, temporal logic

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_tSV9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces FlexTempRule, a neuro-symbolic approach for interpretable time series classification that combines adaptive segmentation, temporal logic rules with different degrees of flexibility, and statistical features. The approach aims to provide model-consistent explanations while maintaining competitive classification performance. Experiments on univariate and multivariate time series.

Critical strengths

The paper addresses a relevant problem, and the gaps in the literature are clearly identified through a well-written and up-to-date introduction. A particularly interesting aspect is the combination of multiple time series representations, following a direction similar to approaches such as HIVE-COTE, MultiRocket, and MR-HYDRA. The integration of flexible temporal rules with statistical features is a promising contribution.

Critical weaknesses

The experimental evaluation is limited to a relatively small number of datasets, which is unusual in time series classification, where extensive benchmarking on the UCR/UEA collections is standard. Moreover, an important interpretable competitor, BORF (Spinnato, F., Guidotti, R., Monreale, A., & Nanni, M. (2024). Fast, interpretable, and deterministic time series classification with a bag-of-receptive-fields. IEEE Access, 12, 137893-137912.), is missing. The comparison should also include traditional baselines such as 1-NN with DTW and state-of-the-art approaches such as ROCKET, HIVE-COTE v2, and MultiRocket-HYDRA. Statistical significance tests are also missing. From an interpretability perspective, examples of extracted rules accompanied by the corresponding time series would strengthen the evaluation. Moreover, I personally found the notation involving squares and diamonds difficult to read. Although established in temporal logic, a more intuitive presentation could improve readability.

Questions for authors

Could the authors extend the benchmarking to a larger collection of UCR/UEA datasets, including BORF and other state-of-the-art competitors, and report statistical significance tests? Could the authors provide concrete examples of extracted rules alongside the corresponding time series in the main paper to better illustrate their interpretability?

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

This paper studies interpretable time series classification. The main goal is to move beyond fixed-patch temporal rule learning and allow temporal rules to have different degrees of flexibility. The proposed method, FlexTempRule, first builds data-driven segments, then learns four kinds of evidence for classification: fixed-position temporal rules, locally slidable rules, globally slidable rules, and simple statistical features. These are combined by an adaptive gating mechanism.

Critical strengths

The paper addresses an important problem. Interpretable time series classification is meaningful, especially in safety-critical applications where users want explicit and understandable decision rules. The paper proposes a reasonably structured extension over prior fixed-patch rule learning approaches. In particular, the distinction between fixed, local-sliding, and global-sliding rules is a sensible and potentially useful way to model temporal uncertainty. The experimental coverage is also fairly broad.

Critical weaknesses
  1. Concerns about rule correctness and interpretability. The supplementary code appears inconsistent with the STL definitions in Eqs. (19)–(20). The nested operators seem to use backward-looking windows, and the Diamond–Box operator checks for an eventual occurrence rather than continuous satisfaction. These differences can change the truth values of the learned rules. It is important to clarify whether the reported experiments used this implementation. I am also not convinced by the interpretability evaluation. The input-dependent MLP gates and classifier biases are retained in Table 5, so high prediction consistency does not necessarily mean that the selected rules faithfully explain the model's decisions.

  2. Insufficient evidence for the main contributions. The ablation study does not directly compare dynamic and fixed patching or control for the increased number of rules and model parameters. The average improvement over TemporalRule is only 0.88 percentage points across eight univariate datasets, with most of the gain coming from ECG200. It remains unclear how much the proposed sliding mechanisms contribute independently.

Questions for authors
  1. Could the authors clarify the discrepancy between the STL definitions in Eqs. (19)–(20) and the supplementary implementation? In particular, the nested operators appear to use different temporal directions and logical conditions. Were the reported experiments conducted using this implementation, and how can the correctness of the learned rules be verified?

  2. Since the final predictions also depend on input-dependent MLP gates and classifier biases, how do the authors establish that the extracted rules faithfully explain the predictions? In Table 5, these components are retained during the feature-retention experiment. Could the authors provide additional ablations to separate their contributions from those of the temporal rules?

  3. Could the authors provide a direct comparison between dynamic and fixed patching, as well as a capacity-matched comparison of sliding and fixed rules? The average improvement over TemporalRule is only 0.88 percentage points, largely driven by ECG200. How can the authors establish that the gains come from the proposed mechanisms and are consistent across datasets rather than being due to increased model capacity or experimental variation?

Reviewer_eK29OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper proposes a method called FlexTempRule for time series classification. It seems that the main novelty the authors claiming are the adaptiveness of their method for both fixed and slidable temporal rules when compared to existing methods.

Critical strengths

The paper seems to have introduced useful extensions.

Critical weaknesses

I can broadly understand the methodological extensions the authors are aiming for (e.g., slidable temporal rules) from the earlier sections. However, starting from Section 3, I found the paper considerably more difficult to follow (with the disclaimer that I am not an expert in time-series classification or temporal logic).

For example, although the operators in Eq. 2 are formally defined, the notation is quite dense, and I found it difficult to understand their roles at an intuitive level. What kinds of temporal patterns are “Always,” “Eventually,” and “Until” intended to capture, and why are these particular operators useful for distinguishing classes of time series? A simple worked example based on one of the motivating applications would help considerably.

More broadly, while the paper formally defines time-series classification and mentions several application domains, I think the motivation could be made much more accessible to readers outside this specific area. It would help to introduce a concrete real-world classification example early on and use it consistently to illustrate the temporal rules and the practical advantage of allowing them to slide.

Questions for authors

N/A

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.