acceptodds
Under review as a conference paper at ICLR 2027

K-ICL: Kernel-based Pair Mining for Instance-wise Contrastive Learning of Time Series

Abstract

Contrastive Learning (CL) is a Self-Supervised Learning (SSL) paradigm that learns representations by attracting *positive* pairs and repelling *negative* ones. SimCLR, a cornerstone CL method, forms positives through label-preserving augmentations of the same instance, and negatives from augmented views of different samples. This instance-wise approach is a key ingredient of many CL solutions, especially in computer vision, but suffers of a major limitation: it is affected by *ill-defined negatives* generated from instances sharing the same semantics. A second limitation specifically concerns time-series, where augmentation can alter the semantic of the instance, producing *ill-defined positives*, which has so far hindered SimCLR in the time-series domain. We address both limitations with K-ICL, a novel probabilistic pair-mining strategy for instance-wise CL that leverages a kernel to measure similarity between time series directly in the input space. Such similarity steers the sampling process towards *well-defined* positive and negative pairs, using only weak augmentations suitable for time-series. Experiments on time-series identification and classification show that K-ICL consistently outperforms SimCLR and its instance-wise extensions, in both few-class and unbalanced settings.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
self-supervised learning, contrastive learning, time series, kernel

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_j7p8OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents K-ICL, a kernel-guided instance-wise contrastive learning framework for time series representation learning. Unlike conventional contrastive learning methods that primarily rely on augmented views of the same instance as positive pairs and treat other instances as negatives, K-ICL uses kernel similarity to guide the probabilistic selection of positive and negative samples. Specifically, samples with higher kernel similarity are more likely to be selected as positives, while those with lower similarity are more likely to be selected as negatives. The authors evaluate K-ICL on multiple time series datasets under both class-balanced and class-imbalanced settings, reporting improvements over existing baselines in clustering and classification-related metrics. They further investigate different kernel functions, sampling strategies, and temperature settings.

Critical strengths

S1. Evaluates the method under both balanced and imbalanced settings.

S2. Includes ablation studies on kernel functions and sampling strategies.

Critical weaknesses

W1. The reliability of kernel similarity as a proxy for semantic similarity is insufficiently investigated. The central assumption of K-ICL is that kernel similarity can provide useful guidance for identifying semantically related instances. However, similarity in the signal space does not necessarily imply semantic or class similarity, particularly for time series with substantial intra-class variations or inter-class similarities. The choice of kernel may also affect how well this assumption holds across different datasets and tasks. Although the paper evaluates different kernel functions, it does not sufficiently investigate the conditions under which kernel similarity reliably reflects semantic relationships. A more comprehensive analysis would help establish the validity and limitations of the central assumption underlying K-ICL.

W2. Probabilistic pair selection may introduce inconsistent training signals across iterations. K-ICL samples positive and negative pairs probabilistically according to kernel similarity. While this design avoids relying on hard thresholds, it also means that the same pair of instances may be selected as a positive pair in one iteration and as a negative pair in another. Such relationship reversals could introduce conflicting gradients and additional optimization noise, especially for pairs whose kernel similarities yield comparable probabilities of being selected as positives or negatives. The current experiments do not sufficiently characterize the extent of this instability or establish whether probabilistic sampling offers a consistent advantage over deterministic alternatives. Although the comparison with ranking-based and threshold-based strategies is useful, a more direct investigation of pair-selection consistency and training stability would strengthen the justification for this design choice.

W3. The manuscript contains several avoidable proofreading and presentation errors. The paper does not appear to have been thoroughly proofread. For example, the caption of Figure 1 describes the two parts of the figure as “Top” and “Bottom,” whereas the actual panels are arranged horizontally, from left to right. Moreover, the figure itself appears to label the proposed method as “KTCL” instead of “K-ICL.” There are also grammatical and spelling errors in the conclusion, including “the learned represenation” and “Future work also aim at”.

Questions for authors

Q1. How reliable is kernel similarity as a proxy for semantic similarity under different augmentation strategies?

Q2. How frequently do pair relationships reverse across training iterations, and do these reversals affect optimization stability?

Q3. What is the specific benefit of probabilistic sampling over deterministic pair selection?

1 comment

Thanks for your review. We address your concerns and questions in the answers below. We made several changes in the manuscript, highlighted in blue in the revised version.

W1 / Q1: kernel similarity as semantic proxy. This is measured directly in the paper: pair precision and recall (Table 1) quantify how well kernel similarity reflects class membership, since if it were a poor proxy, high-similarity pairs would often be different-class and positive precision would be low. We now state this explicitly in Sec. 4.2. Its dependence on the kernel is analyzed in the kernel ablation (Table 8): the assumption holds strongly for alignment-aware kernels (GAK, DTW) and degrades for Euclidean. We do not assume signal similarity guarantees class agreement. It can fail under high intra-class variation or inter-class similarity, which is why Sec. 3.3 / App. B make the guarantee conditional on similarity-score separation, with an overlap-aware extension.

The paper thus provides empirical evidence of both the usefulness and the limitations of the proxy on the evaluated datasets.

W2 / Q2, Q3: sampling consistency and stability. We measured this directly in two new experiments (App. A.7, Figs. 6-7):

  1. Resampling fixed batches 2,000 times; only 1-3% of pairs ever reverse, concentrated on ambiguous mid-similarity pairs, which are mostly not selected. The most similar 10% reverse in under 0.12% of cases, and the anchor's own view stays positive 98.7-99.9% of the time. Since the kernel is computed on the inputs, these probabilities are fixed by the data and do not drift during training.

  2. At fixed weights, resampling only the pairs perturbs the gradient far less than resampling a full mini-batch step, so pair sampling adds less noise than the stochasticity CL already tolerates. For ambiguous pairs no correct hard label exists in our unsupervised setting, so soft exposure is preferable to a possibly wrong hard assignment (as in hard vs soft pseudo-labeling). This is consistent with our sampling-rule ablation (Table 5), where probabilistic sampling is the most robust.

W3: proofreading. Fixed.

Reviewer_4gXeOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript proposes K-ICL, which utilises kernel similarity between time series to screen pairs of positive and negative samples, thereby reducing unreasonable pairings and improving the representations obtained through contrastive learning.

Critical strengths

The method is designed in a relatively straightforward manner, utilising kernel similarity to guide the identification of positive and negative samples, and is easy to understand.

Critical weaknesses

Major issues

1.SimCLR is a classic contrastive learning framework, and several methods have subsequently been developed to address its potential shortcomings. This manuscript should begin by outlining the strengths and weaknesses of existing time-series contrastive learning algorithms, and use this as a basis to explain its motivation.

2.Ill-defined negatives and positives are a classic problem in the field of contrastive learning. How does the solution proposed by the authors differ from other existing methods?

3.The description of the main contributions is not sufficiently clear. The authors have not explained separately the mechanisms by which K-ICL reduces the number of positive samples with semantic inconsistencies, nor how it reduces the incidence of similar samples being misclassified as negative samples.

4.Much of the relevant work draws on research in the field of computer vision. Although some time-series methods have been addressed, the analysis of their technical differences and limitations remains insufficient.

5.The question statement lacks a formula of the form y = f(x).

6.The number of datasets used is relatively small, and the sample size of each dataset is also rather limited; it is recommended that representative comparative methods from 2025 onwards be included.

7.Although kernel similarity computations do not participate in backpropagation, pairwise computations within a batch may still incur significant overhead. It is recommended to report training time, time per epoch, and peak memory or GPU memory usage, and to analyse how these overheads vary with batch size, sequence length and data scale.

8.The proposed algorithm is essentially a data augmentation method; it will be compared with this and other time series data augmentation methods to validate the effectiveness of the proposed method.

9.There is a lack of illustrative examples and visualised experiments to demonstrate that the proposed algorithm is effective.

10.In the introduction, the authors state that the proposed method ‘integrates seamlessly into instance-wise CL pipelines for time series’, but no corresponding experiments are provided. We recommend incorporating K-ICL into several existing contrastive learning frameworks (such as MoCo, BYOL, SimSame, etc.) as a plug-in or a loss function to validate its effectiveness.

Minor issues

1.Figure 1 uses a left-right layout; the terms ‘Top’ and ‘Bottom’ in the figure caption should be changed to ‘Left’ and ‘Right’ respectively. Furthermore, what does ‘KTCL’ mean in the figure?

2.In Tables 2 and 3, some methods are identified solely by author and year, whilst others are identified by name; this inconsistency in presentation makes it difficult to identify them quickly.

3.The format of the references is inconsistent.

Questions for authors

1.Compared with similarity-weighted and sample-pair mining methods in existing contrastive learning for time series, what is the core innovation of K-ICL? Why is probabilistic sampling adopted rather than direct weighting or deterministic filtering?

2.How does K-ICL control errors introduced by cross-instance positive samples and reduce false negatives? Could you illustrate the respective contributions and limitations of positive and negative sample mining through ablation experiments and specific case studies, where each is enabled separately?

3.Does K-ICL still achieve stable gains when integrated into other frameworks with compatible instance-level comparison objectives? Compared to the original framework, by how much does kernel computation increase training time and memory overhead?

4 comments

Thanks for your review. We address your concerns and questions in the answers below. We made several changes in the manuscript, highlighted in blue in the revised version.

W1: motivation from existing TS-CL. Section 2 already discusses advances in time-series CL, including augmentation design, temporal objectives such as TS2Vec and TS-TCC, and input-similarity weighting such as SoftCLT. Our motivation concerns the quality of the instance pairs used for learning: richer temporal objectives do not, by themselves, prevent same-class instances from being repelled, while stronger augmentations can alter time-series semantics. K-ICL addresses this problem by recovering informative positives and avoiding likely false negatives using input-space similarity. We stressed more in the introduction this connection to the existing time-series literature.

W2: difference from existing methods. The differences from existing methods are already discussed in detail in Section 2, and the relevant alternatives are compared experimentally in Tables 2–3. Our distinctive mechanism, defined in Section 3.2, uses input-space kernel similarity to probabilistically select both positive and negative pairs. Unlike representation-based mining, this selection does not depend on the quality of the evolving embeddings.

Section 2 also explicitly discusses the closest input-space alternative, SoftCLT: it weights pair contributions, whereas K-ICL samples which pairs enter the contrastive objective. Section 4.1 explains how we evaluate its instance-wise component, SoftCLT-I, under the same encoder, augmentations, and training protocol. The paper therefore already provides both an explicit account of the methodological differences and a controlled empirical comparison against the relevant alternatives.

W3: separate the two mechanisms. The two mechanisms are already distinguished in Section 3.2: positive mining selects candidates with higher kernel similarity to the anchor, including from other time series, adding diversity while keeping augmentations weak; negative mining preferentially samples lower-similarity candidates, making similar instances less likely to be wrongly repelled (selected positives excluded from the negative pool).

To quantify each contribution separately, we added a positive-vs-negative mining ablation (Sec. 4.3, Tables 6 and 13). Positive mining alone adds cross-instance positives and already lowers the FN rate (32%->19-23% unbalanced) and negative mining alone drives it lowest (3-6%), but selects the easiest negatives, not always raising accuracy. The two are complementary and every variant stays above SimCLR.

W4: CV vs TS work. The false-negative problem arises from instance-wise pair construction and is shared across data modalities. Many established remedies originated in computer vision, but their debiasing, weighting, and pair-selection mechanisms operate on the contrastive objective. Their relevance to time series is precisely what our experiments assess, using the same time-series encoder and weak augmentations across methods. The time-series literature is also explicitly analyzed in Section 2. Our selection thus covers both general algorithmic solutions to false negatives and the closest time-series-specific alternative. Restricting the discussion to methods originally introduced for time series would omit directly relevant solutions to the problem our paper addresses.

W5: y=f(x) Our setting is two-stage self-supervised representation learning and does not reduce to a single supervised y=f(x). Section 3.1 defines the representation mapping h=E(S). We now make its connection to downstream prediction explicit, h=E(S) and y_hat=g(h). The two components are trained separately: the encoder E learns representations from unlabeled time series using the contrastive objective (our focus); then E is frozen and a classifier g is fitted on a small labeled set. Labels supervise only the downstream classifier, not the representation learning addressed by K-ICL. We added this two-stage formulation to Sec. 3.1 of the revised paper.

W6: datasets / recent methods. Our experiments already cover all the core claims, measuring both pair quality and downstream performance on five datasets spanning different modalities and class counts, under balanced and imbalanced conditions, with a shared pipeline and ten matched runs. This directly tests both the proposed mechanism and its practical benefit across the conditions motivating K-ICL. Recent methods are already considered in the paper: Section 2 discusses AdaTS and SIP-LDL (2025). Their omission from the numerical comparison has specific reasons: the official repo of AdaTS is empty, while SIP-LDL uses labels during representation learning, changing our fully unsupervised setting. We include the directly relevant, reproducible SoftCLT-I under the same encoder and training protocol.

W7: cost / scaling. We added training time, per-epoch time and memory in Table 12 (App. A.6). Since the kernel lies outside the computational graph, it adds no backpropagation memory. Naive per-batch GAK is the only overhead: because it depends solely on the inputs, caching reduces it to 1.3-1.5x SimCLR's training time, and precomputing all similarities once before training removes any kernel computation from the training epochs entirely. All of this is confined to training: at inference the kernel is never used, so K-ICL matches SimCLR exactly in latency and memory. Finally, GAK cost grows with sequence length, but our inputs are short and semantically meaningful windows (e.g. two to three heartbeats for ECG), so the cost stays low.

W8: "it is a data augmentation method". K-ICL introduces a pair-mining strategy, not a new augmentation operator. Augmentation determines how a time series is transformed into a view; K-ICL determines which candidate views should form positive or negative pairs with an anchor. The proposed mining step does not generate or modify time series.

This separation is explicit in Figure 2: Step 1 constructs views using existing weak augmentations, while Steps 2-4 compute similarities and select pairs. All compared methods use the same augmentation protocol (Section 3.2 and Appendix A.2), so the reported gains cannot be attributed to a different augmentation recipe.

Our comparisons therefore hold augmentation fixed and vary the pair-mining or weighting strategy. This directly evaluates the proposed contribution against relevant alternatives (Tables 2–3). Augmentation design concerns a separate component of the pipeline, which K-ICL leaves unchanged.

W9: visualizations. We added t-SNE embedding projections: Fig. 3 (CEBS, main) and Fig. 5 (PenDigits, App. A.5), where K-ICL forms more compact, better-separated clusters than SimCLR, consistent with its clustering metrics. These are illustrative, with ARI/NMI as the quantitative evidence.

W10: integration into other frameworks. Our experiments already demonstrate integration into a SimCLR-based pipeline. As stated in the Introduction, K-ICL is a principled extension of SimCLR. We replace its pair construction while retaining the encoder and augmentation protocol, and compare the resulting representations against SimCLR and alternative mining strategies.

The suggested frameworks also require different considerations. BYOL/SimSiam use no negatives, so incorporating K-ICL’s full positive-and-negative objective would change their learning objective. MoCo is a plausible extension, but adapting mining to its queue-based training would constitute an additional framework-specific study. We revised the Introduction to state that K-ICL integrates into a SimCLR-based instance-wise pipeline by replacing its pair construction.

Q1: core innovation / why probabilistic. The distinction is already explicit in Section 2: “using kernel similarity to guide explicit probabilistic pair sampling.” K-ICL selects both positive and negative pairs using input-space similarity, independently of the evolving embeddings. Unlike SoftCLT-I’s direct weighting, it determines which pairs enter each update, rather than assigning weights to their contributions.

Probabilistic sampling favors high-similarity positives and low-similarity negatives without requiring a fixed similarity cutoff or a predetermined positive fraction. It also allows different candidates to contribute across updates, balancing selective pair construction with broader coverage. These choices are already tested experimentally. Tables 2-3 compare against direct weighting through SoftCLT-I. Section 4.3 and Table 5 replace probabilistic sampling with threshold-based and rank-based selection while keeping the kernel fixed. Table 8 shows that K-ICL maintains negative precision above 90% on all five datasets and achieves the highest negative precision on four. Thus, the sampling choice has both an explicit rationale and direct evidence supporting its ability to avoid false negatives.

Q2, Q3. See W3 (positive/negative ablation) and W7 (cost).

I appreciate the additional ablation experiments evaluating positive and negative pair mining separately. These results clarify the contribution of each mechanism. I have decided to maintain my original score.

I would appreciate clarification on one discrepancy between W3 and Table 6. W3 reports a false-negative rate of 3–6% for negative-only mining, while Table 6 gives a range of 3–8% for negative-only mining and 3–6% for the full K-ICL method. Please also report the false-negative rates in Table 6 as percentages rounded to two decimal places.

Thank you for flagging the inconsistency, which we have promptly fixed in the revised manuscript.

However, we are concerned that your follow-up leaves the basis for the unchanged recommendation unclear. Our rebuttal responded individually to every weakness and question you raised, including the motivation and novelty, choice of comparisons, computational cost, and distinction between pair mining and augmentation. Beyond discussing the additional ablations and the numerical inconsistency, your follow-up does not explain which of these responses remain insufficient.

We ask you to identify which concerns have been resolved, which remain, and why the remaining concerns still justify rejection. If an answer or revision is considered inadequate, please state the specific reason so that we can respond. Without this explanation, the discussion leaves unclear how our responses and revisions have been taken into account.

The ICLR reviewer guidelines (https://iclr.cc/Conferences/2027/ReviewerGuidelines) explicitly ask reviewers to "State your reasoning and what did/didn't change your recommendation throughout the discussion phase." We ask that this reasoning be made explicit here.

With the discussion period closing, we also ask the Area Chair to examine the recommendation in light of our full response and revised manuscript, and to ensure that the final assessment clearly identifies any substantive concerns that remain after the rebuttal.

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

This paper proposes K-ICL, a kernel-guided pair-mining strategy for instance-wise contrastive learning on time series. Their approach aims to reduce false negatives resulting from instances sharing the same semantics within the batch and from reliance on strong domain-dependent augmentation in time series. They achieve pair mining by measuring similarity between time series directly in the input space using the Global Alignment Kernel (GAK). The authors claim that K-ICL reduces the number of ill-defined negatives, especially in imbalanced-class and few-class regimes.

Critical strengths

The paper is well-written and easy to follow. Particularly, the introduction clearly delineates the problem of Instance-wise CL in time series while citing the appropriate literature, and subsequently properly positions this work and how it relates to the seminal SimCLR paper from computer vision. The notations are also well-defined and make following the methodology section easier. Also, adding clustering metrics in the evaluation, such as the NMI and ARI, is a good way to measure representation quality beyond linear probe accuracy.

The provided code is useful and provides additional information about the K-ICL method and training process.

Critical weaknesses

Although I have no major soundness concerns regarding the proposed method, using GAK to hard-mine positives and negatives is straightforward and intuitive. Nevertheless, there are a few points that need clarification. First, using the GAK kernel to compute similarities is prohibitively slow and computationally expensive, and is not feasible, especially for long sequences, which are common in time series. To warrant the use of this expensive positive and negative mining step in the vanilla InfoNCE objective, the gain has to be worth it, and so far, the experiments on just 3 out of the 128 UCR benchmarks have not done a good job of showing that.

Regarding the statement in lines 271–272, "differently from SimCLR, we can sample multiple positive pairs per anchor, gathering additional information that is beneficial for CL training":

Gathering multiple positive pairs per anchor has never been the issue; the problem remains how to integrate this into a smooth, parallelizable training procedure. Specifically, SimCLR can create several potential augmented views for a single anchor, while a single positive per anchor admits a clean and fully parallelizable matrix implementation via the similarity matrix in InfoNCE; extending this to multiple positives per anchor breaks the single-label cross-entropy formulation.

In the introduction, a major motivation for hard negative mining using the expensive GAK is to avoid augmentations that risk distorting the time series signal. Nevertheless, in line 194, the supposedly "weak" augmentation contains the typical time series data-distorting augmentations, specifically scaling and jittering. It would be nice to see how the K-ICL method performs without augmentations at all.

Moving on to the experiments, the baselines and dataset coverage are not convincing. Although the authors claim that they excluded methods centered on augmentation design or training that leverages temporal structure, like TS2Vec and TS-TCC, because they introduce additional components to this pipeline, this is not correct. Both method pipelines are designed to integrate seamlessly with any instance-wise CL pipeline. Specifically, TS2Vec (Figure 1) shows that the context views are generated by applying timestamp masking and random cropping to a single instance of a time series. Also, if the TS2Vec pipeline is incompatible with the K-ICL pipeline, then SoftCLT should not be either, as SoftCLT is not a standalone architecture and is built on top of TS2Vec (Table 2, SoftCLT).

Lastly, evaluating K-ICL on the entire UCR benchmark suite would help support the contribution of this method and ascertain whether the extra computational cost of hard positive mining is worth it.

Questions for authors

Most of the questions are elaborated in the weakness above.

  1. How do you intend to integrate multiple positive pairs per anchor in the InfoNCE loss?

  2. In the appendix, it shows that K-ICL was trained for 500 epochs at a batch size of 64; this is significantly more than the usual training budget in the time series literature for the UCR benchmark. How was the number of training epochs decided?

  3. In your code, you computed the GAK inside the training loop (for 500 epochs). Did you compare the training time with other baselines?

  4. Did you try contrasting SaS_{a} and SjS_{j} directly without the handcrafted augmentations?

Minor typos:

Lines 062- 063: the bolded Figure 1 captions should be Left and Right, not Top and Bottom.

3 comments

We thank the reviewer for the positive assessment of the writing, positioning, notation, and clustering metrics. We address each point; the revised text is marked in blue in the paper.

W1/Q3: GAK cost and benefits. We added measurements in the revised paper (Table 12, App. A.6). It is true that the kernel introduces additional computation, but this is confined entirely to training: GAK only selects the training pairs, and for everything else K-ICL is identical to SimCLR. At inference, K-ICL is simply a trained encoder, without kernel or extra parameters, so its latency, memory and model size match SimCLR exactly. During training, the kernel lies outside the computational graph, adding no backpropagation memory. Naive per-batch GAK is the main overhead, but since the kernel depends only on the inputs, it can be cached: caching brings K-ICL to 1.3-1.5x SimCLR's time, and precomputing once removes it from the loop entirely. The mining step costs at most 0.015s/epoch. GAK cost scales with sequence length, but our inputs are short and semantically meaningful windows (e.g. two to three heartbeats for ECG) rather than long raw signals, so this cost remains low. The windowing reflects the natural unit of the task rather than a constraint imposed by the kernel.

Our proposed method has concrete benefits that justify this moderate, training-only cost. Under imbalance, the proportion of incorrect negatives falls from approximately 26–38% to 5–10% across the five datasets. On PAMAP, for example, MLP accuracy improves from 78% to 95% over SimCLR (Tables 1 and 3).

W2/Q1: multiple positives. We believe there is a misunderstanding. Eq. 3 already sums the positive terms in the numerator and the positives+negatives in the denominator, both computed in a single log-sum-exp over the chosen entries, regardless of their number. This is a standard multi-positive contrastive objective, as in supervised contrastive learning (Khosla et al., 2020), evaluated on the similarity matrix with masks, and it need not use single-label cross-entropy. We clarified this point in the revised Sec. 3.2.

W3/Q4: weak / no augmentation. K-ICL is a principled extension of SimCLR, which is built on augmentation, so we keep SimCLR's view to isolate the mining effect, and our augmentations are minimal (App. A.2). A view distorting the signal would get low kernel similarity and would not be mined as a positive. We added an ablation removing the augmented view entirely (App. A.9, Table 14): positives/negatives are mined among the original series only, and ARI changes by at most 0.02 while K-ICL stays far above SimCLR. Removing augmentation thus does not reduce K-ICL's advantage: the slightly larger drops under imbalance are expected, as every positive must then be another series and minority classes have few same-class partners.

W4: TS2Vec/TS-TCC/SoftCLT. We never claimed incompatibility: our Related Work states these methods retain SimCLR's instance-wise InfoNCE loss, so our mining remains relevant to them. Our comparison is designed to establish whether better pair construction improves representation learning, so we compare K-ICL against competing mining and weighting strategies using the same encoder, augmentations and training protocol: holding these components fixed lets us attribute the observed gains to the pair-construction strategy.

The masking and cropping the reviewer notes in TS2Vec's describe its view-generation procedure, but TS2Vec additionally introduces temporal contrast at multiple resolutions. Comparing complete pipelines would therefore change several components simultaneously and would not isolate the contribution of pair mining.

SoftCLT is treated by exactly the same criterion, which resolves the apparent inconsistency the reviewer raises: we do not compare against the complete TS2Vec+SoftCLT framework either, but against its instance-wise soft-assignment component, SoftCLT-I, within our shared pipeline (Sec. 4.1, Tables 2-3). SoftCLT-I is in fact a direct competitor to our contribution: both use input-space similarity to guide instance-wise learning, through pair weighting in SoftCLT-I and probabilistic pair selection in K-ICL. For example, on imbalanced PAMAP, K-ICL improves MLP accuracy from 83% with SoftCLT-I to 95% (Table 3). Together with the reduction in false negatives (Table 1), this supports our central claim that improving pair construction yields better representations under controlled conditions. We rephrased the relevant sentences to make this comparison criterion clearer.

W5: dataset coverage. Our evaluation targets the specific problem K-ICL addresses: unreliable pair construction under few/imbalanced classes. We test these conditions across five datasets spanning different modalities and class counts, with ten runs and a fixed pipeline, measuring both the quality of mined pairs and downstream representation quality. These datasets already make the target property evident, under both balanced and imbalanced regimes, so the proposed mechanism and its benefit are clearly established. Adding further datasets where the same conditions hold would not change this conclusion. Evaluating the entire UCR archive is therefore not a prerequisite for our contribution.

Computational value is a separate question, depending on training cost relative to performance gains rather than on the number of datasets and we address it above. K-ICL's extra kernel computation is confined to training, while inference uses the encoder alone.

Q2: 500 epochs. 500 epochs is a protocol fixed across all methods (same epochs, batch size 32, optimizer, lr, splits), with no early stopping or checkpoint selection, so no method gains implicit model selection. The value is a choice, not a claim of necessity or optimality.

Typo. Fixed.

Reviewer_VAFfOpenReview ↗

Response to Authors Rebuttal

Thank you for your response and for adjusting the paper to address some of the recommendations. The SoftCLT-I variant you compared against is clear to me now.

GAK cost scales with sequence length, but our inputs are short and semantically meaningful windows

This reinforces the fact that K-ICL will struggle on the entire UEA or UCR benchmark, where some datasets have sequence lengths in the thousands.

Adding further datasets where the same conditions hold would not change this conclusion

I disagree with this argument; having good performance in a small subset of the UCR is not representative of the entire benchmark suite.

Again, evaluating on a few datasets from the UCR does not support the contribution of this method or ascertain whether the extra computational cost of hard positive mining is worth it.

Comments on the paper

1 comment
Area_Chair_WSgbOpenReview ↗

Reminder

Dear Reviewers,

Thank you for your efforts so far. This is a friendly reminder to review the authors' rebuttal and respond to their comments as soon as possible, before the discussion deadline (November 18 AoE). Please also indicate whether your concerns have been addressed. Thank you for your time and cooperation!

Best regards, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.