acceptodds
Under review as a conference paper at ICLR 2027

TEX-MEX: Similarity-Guided Temporal Contrastive Learning for Time Series

Abstract

Contrastive learning for time series depends on deciding which pairs of inputs should be treated as similar and dissimilar. Because few augmentations reliably preserve meaning in time series, prior methods have relied on two disjoint notions of similarity between windows: structural similarity, which relates windows with similar morphology, and temporal similarity, which relates windows that occur close together in time. We introduce TEX-MEX, to our knowledge the first approach to unify the two into a single relationship. Given a random batch of windows drawn across the dataset, TEX-MEX uses the encoder to identify which windows are structurally similar or dissimilar to each other. It then pulls each window toward the temporally shifted versions of similar windows and pushes it away from those of dissimilar ones, learning what typically accompanies similar windows in time. Across PPG, EEG, ECG, accelerometry, and continuous glucose monitoring, TEX-MEX broadly achieves state-of-the-art performance against contrastive, reconstructive, JEPA-based, and modality-specific approaches, establishing it as a strong general SSL objective. We further train a 560K-parameter PPG foundation model that achieves state-of-the-art performance across a broad range of downstream benchmarks while being roughly 50× smaller than the next-best unimodally pretrained PPG foundation model in our evaluation. We will release this foundation model open-source and open-weight to the community to support future work and practical applications.

Primary area:
learning on time series and dynamical systems
Keywords:
self-supervised learning, foundation models, timeseries, biosignals

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_U2kwOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes TEX-MEX, a method for integrating structural and temporal information into the contrastive objective. It does this by extracting semantically similar patterns from time series using the encoder's output features and combining them with temporal similarity, which draws positive and negative pairs from the temporal context of similar and dissimilar windows. The authors validate this by evaluating linear probe classification performance on physiological time series signals such as PPG, ECG, and HAR datasets. In addition, they extend their pretraining recipe to a unimodally pretrained PPG foundation model and compare it with general time-series as well as PPG-specific foundation models.

Critical strengths

The paper presentation is generally understandable. The tables and figures are legible and easy to follow to an extent. Also, attempting to combine semantic and temporal information in the contrastive objective while avoiding traditional augmentations that risk distorting time series signal is desirable in time series representation learning.

Critical weaknesses

The contrastive objective needs significantly more explanation of the learning signal and the choices for positive selection. Specifically, the initial similarity scores between the samples in the batch and the anchor are computed from representations produced by an untrained encoder. Why should these random similarity priors guide the entire learning process, and what incentive does the encoder have to learn discriminative features as training progresses? Also, the selection of random past and future windows assumes non-IID data. This is not the case for most time-series representation learning methods for classification, which assume IID data and shuffled batches, with all the information needed for learning contained within a single instance of a predefined window size (UCR and UEA). I understand that TNC does not follow this convention, but most recent methods do (TS2Vec, SoftCLT, TFC, TSTCC).

Also, the components of the loss function do not appear to be well motivated. The w+w^+ term, which assigns a lower weight to the dissimilar sample, should be sufficient as a soft contrastive objective; adding an NN component with a reverse weight that goes into the denominator is redundant at best. In addition, the statement in lines 204–205 that "P and N do not require representations to preserve features that distinguish a window from its temporal shift" is unsubstantiated, as w+w^+ and w−w^- in Equation 1 already have anti-collapse terms in the denominator, which are usually sufficient to prevent representation collapse in contrastive learning.

Given that the loss function contains several redundant components, which would inadvertently increase training time, it would be interesting to see ablations on a more comprehensive classification benchmark suite, such as the UCR and UEA benchmarks, rather than on the more niche CGM metabolic subphenotyping task.

Overall, this paper's main contribution appears to be a substitution of the DTW similarity in SoftCLT with similarity computed using a randomly initialized encoder. For IID data, the proposed shift is not practical, and the initial similarity scores from a randomly initialized encoder backbone provide a poor prior. Therefore, a clear positive and negative selection mechanism to guide the contrastive objective is important. Also, it is important to clarify that SoftCLT is not a standalone architecture but rather is applied to other contrastive methods with a clear learning signal, such as TS2Vec.

TS-TCC - Eldele, E., et al. "Time-Series Representation Learning via Temporal and Contextual Contrasting." IJCAI, 2021.
TFC - Zhang, X., et al. "Self-Supervised Contrastive Pre-Training for Time Series via Time-Frequency Consistency." NeurIPS, 2022.

Questions for authors

Can the author clarify the following points?

  1. Is the encoder backbone a pretrained backbone or randomly initialized?

  2. How are the datasets processed? assuming iid with shuffle True, or keeping shuffle False as in non-iid?

  3. In designing the loss components in eqn 2, 3, and 4, did you use any parallelization techniques in your code?

  4. What is the main invariance your method aims to learn?

  5. Can you elaborate on the main motivation for including the extra H term in the loss function in eqn 4?

Minor issue:

Lines 206 - 208 mention contrasting the "temporal shifts" of similar batch members against their corresponding unshifted windows; however, the H component doesn't factor in any temporal shifts

7 comments
AuthorsOpenReview ↗

Response to Reviewer U2kw (1/5)

We thank the reviewer for their detailed and thoughtful feedback. Their questions have helped us clarify several aspects of TEX-MEX's motivation, learning dynamics, and loss design. We have revised the manuscript and added targeted ablations and runtime measurements in response. In brief:

  • Distinction from SoftCLT. SoftCLT learns cross-recording similarity and temporal proximity separately, with temporal supervision restricted to each recording's own observations. TEX-MEX combines these relationships to learn from how similar patterns evolve across different recordings, capturing a richer diversity of temporal contexts than any single recording can provide (§1).
  • Learning from random initialization. Each anchor's own temporal shift guarantees a positive signal from the first training step. Random convolutional weights have also been used successfully as features in prior work, since the same fixed transformation is applied to every window. The encoder-based similarity measure then improves as training progresses (§2).
  • Data requirements. TEX-MEX draws on modest context beyond a single window, requiring only recordings long enough to retrieve a nearby past or future window. Such continuous recordings are also a common type of time series data for representation learning, and sufficiently long pre-segmented instances can be rewindowed (§3).
  • Roles of N and H. N supplies explicit negative supervision and reduces false negatives through similarity-based grading. H is designed to discourage shortcuts based on persistent, spurious features that remain unchanged across temporal shifts. Our ablations support the value of both terms, including a new uniform-negative-weight experiment motivated by this review (§4).
  • Efficiency and evaluation. New runtime measurements show negligible overhead from N's graded weighting and H. We also clarify the scope of our component ablations and the applicability of UCR/UEA-style datasets (§5).

We address each concern below.

1. Relation to SoftCLT

"Overall, this paper's main contribution appears to be a substitution of the DTW similarity in SoftCLT with similarity computed using a randomly initialized encoder."

SoftCLT models cross-recording similarity and temporal proximity through separate losses, whereas TEX-MEX combines them within the same contrastive pair, allowing each window to learn from the varied temporal evolutions of similar patterns across recordings rather than only its own observed temporal context. Specifically, SoftCLT's instance-wise loss relates different series at the same timestamp, while its temporal loss relates nearby timestamps within a single input (Lee et al., 2024). Neither constructs pairs linking related windows across both recordings and time.

TEX-MEX instead grades similarity between source windows and uses their temporally shifted counterparts as contrastive examples. A single recording provides a limited view of how a pattern evolves, whereas similar patterns across recordings may occur in a much broader range of temporal contexts. By composing these relationships, TEX-MEX can learn from this diversity rather than restricting temporal supervision to one observed trajectory. We discuss this distinction further in the General Response ("What distinguishes TEX-MEX?").

The advantage of this distinction is supported empirically. TEX-MEX outperforms SoftCLT on 18 of 20 results, by up to 14.1 AUROC (Tables 1-2). Removing the similarity-based grading that enables cross-recording temporal-context contributions also degrades 11 of 12 metrics, by up to 17.1 AUROC (Table 3; also shown in the General Response).

TEX-MEX also makes learned, adaptive similarity grading both effective and highly efficient. SoftCLT reports that applying embedding-space contrastive approaches such as NNCLR and ASCL to TS2Vec reduced performance (Lee et al., 2024, Sec. 4.6). In TEX-MEX, however, the encoder's similarity measure improves during training and ultimately surpasses DTW and REBAR at predicting shared class membership (Fig. 3). In our new grader ablations, substituting fixed DTW or REBAR similarity within TEX-MEX degrades 11 and 10 of 12 metrics, by up to 7.8 and 7.7 AUROC, respectively (revised Table 3). The encoder-based grader also reuses representations already computed in the forward pass, whereas per-batch DTW and REBAR grading increase training time per step by 5.9x and 13x. SoftCLT's approach of precomputing pairwise DTW similarities requires a quadratic number of preprocessing comparisons in the number of training inputs, an overhead TEX-MEX avoids.

(continued in part 2/5)

AuthorsOpenReview ↗

Response to Reviewer U2kw (2/5)

(§1, continued)

"SoftCLT is not a standalone architecture but rather is applied to other contrastive methods with a clear learning signal, such as TS2Vec."

Our SoftCLT comparison uses TS2Vec, following the original paper's default setup. We have clarified this in the revised manuscript (Section 4.1, line 267).

2. Learning signal from a randomly initialized encoder

"Why should these random similarity priors guide the entire learning process, and what incentive does the encoder have to learn discriminative features as training progresses?"

The encoder is indeed randomly initialized rather than pretrained. Because the similarity weights are recomputed at every step from the current encoder, they become increasingly informative as training progresses (Fig. 3), and two properties make the random starting point workable.

First, a valid positive signal is guaranteed from the first training step. Each anchor's own temporal shift always contributes to the positive term P because its source-window self-similarity is cii=1c_{ii}=1, regardless of initialization. Training therefore does not depend entirely on accurately identifying similarities between different recordings before any learning has occurred. This provides a reliable starting signal, although the substantial performance drop without graded weighting (Table 3) shows that the self-shift alone is insufficient to achieve TEX-MEX's full performance.

Second, we respectfully note that random convolutional features have been shown to carry useful structure. At initialization, the encoder applies the same fixed transformation to every window, so similar input patterns tend to receive similar representations. Random convolutional architectures have been shown to preserve useful structural properties before training (Saxe et al., 2011), and in time series, ROCKET achieved strong classification performance on the UCR archive using random convolutional features (Dempster et al., 2020). In this way, the randomly initialized encoder can therefore offer a reasonable starting point for similarity estimation, which TEX-MEX then improves as training progresses (Fig. 3). Our new grader ablations are consistent with this: although the encoder's similarity starts from random initialization, replacing it with fixed DTW or REBAR distances, which are available from the first step, degrades 11 and 10 of 12 metrics, by up to 7.8 and 7.7 AUROC, respectively (revised Table 3).

The incentive to learn discriminative features comes from the contrastive objective itself. It rewards representations that distinguish the temporal contexts associated with similar source windows from those associated with dissimilar ones. As the encoder learns these distinctions, its similarity estimates become more informative, sharpening the weights used in subsequent training. Figure 3 evaluates this learned similarity directly as a distance function for identifying windows that share a class: after one epoch, it is within approximately 2 percentage points of REBAR, and by mid-training it exceeds REBAR by 4-5 points.

These observations support the viability of learning from random initialization. TEX-MEX begins with a reliable positive signal and input-dependent similarity estimates, its grading function becomes more informative during training, and the resulting representations achieve strong downstream performance. The graded-weighting ablation further demonstrates that this similarity-based supervision is important to TEX-MEX's effectiveness (Table 3). We have clarified these points in the Methods section of the paper (Section 3, lines 196–201).

(continued in part 3/5)

AuthorsOpenReview ↗

Response to Reviewer U2kw (3/5)

3. Data assumptions

"Also, the selection of random past and future windows assumes non-IID data… For IID data, the proposed shift is not practical."

TEX-MEX does use context beyond a single window, but this requirement is modest, and the continuous recordings it relies on are also a common type of time series data for representation learning. At each step, we randomly sample a batch of windows across the pretraining dataset, as in standard SSL, and pair each with a past or future window from its own recording, using a single temporal offset shared across the batch (Section 3). Our PULSE comparison, for example, retains the same source-window sampling procedure, with TEX-MEX additionally retrieving each shifted counterpart.

Specifically, TEX-MEX needs only recordings long enough to provide a shifted partner, typically a short distance into the past or future. As the reviewer notes, established general-purpose SSL methods such as TNC and CPC construct temporal supervision from continuous recordings, and more recent works such as REBAR and RelCon draw comparisons from windows across a recording. Large-scale models such as MOMENT and Chronos also harness extended sequences during pretraining. Further, even where only pre-segmented sequences are available, sufficiently long ones can be divided into shorter windows, as in our CGM experiments (Section 4.2).

We agree that short, isolated sequences, such as those in some UCR/UEA datasets, may be less suited to TEX-MEX pretraining when subdividing them leaves too little signal structure for meaningful cross-recording similarity, and we now note this limitation in the paper's conclusion (Section 6, lines 521–523). At the same time, long, continuous recordings are a widespread form of time series data, arising in wearable sensing, clinical monitoring, and many other domains where signals are recorded over extended periods. TEX-MEX targets continuous biosignals in general and is especially well suited to mHealth and wearable data, and we have scoped our claims accordingly (Section 6, lines 521–526). Within this setting, TEX-MEX broadly achieves state-of-the-art performance across five signal domains (Tables 1-2, 4).

Finally, this requirement applies only during pretraining: the pretrained encoder processes downstream windows independently, without their surrounding temporal context.

4. Motivation for N and H

"The term, which assigns a lower weight to the dissimilar sample, should be sufficient as a soft contrastive objective; adding an component with a reverse weight that goes into the denominator is redundant at best."

This question highlights an important distinction between TEX-MEX's loss and SoftCLT's. SoftCLT can rely on soft assignments alone because its targets are fixed and external, whereas TEX-MEX's are learned. SoftCLT trains the model's softmax similarities to match soft assignments computed from DTW distances on the raw data (Lee et al., 2024). Because these targets do not depend on the encoder, a collapsed representation cannot satisfy them.

In TEX-MEX, the similarity weights are derived from the encoder being trained, so an objective built only from them could be trivially satisfied by collapse: if all representations coincide, the weights become uniform and agree with the uniform similarities. N prevents this by explicitly placing the temporal contexts of dissimilar source windows in the denominator, so that P must attract the temporal contexts of similar windows relative to dissimilar ones.

N's similarity-based grading is also important. Uniformly treating other windows as negatives risks repelling semantically related examples, whereas N assigns lower negative weight to windows that appear similar to the anchor. This robustness to false negatives is a motivation shared by prior soft contrastive approaches, including RelCon and SoftCLT (Xu et al., 2025; Lee et al., 2024). We have added an additional ablation replacing N's graded weights with uniform negative weights. This degrades 10 of 12 metrics, by up to 6.6 AUROC (revised Table 3). The experiment supports the value of similarity-dependent negative grading rather than treating all candidate negatives equally. N therefore provides complementary negative supervision, while its grading helps reduce the effect of false negatives.

(continued in part 4/5)

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

The paper proposes a new temporal contrastive learning for pre-training model. This method called Tex Mex, allow to align different windows by structure and time. This method is compared on different modalities and show strong performances.

Critical strengths
  • The motivation behind the paper is very clear and the method answer to that by unifying structural and temporal similarity.
  • The results are strong over different modalities that shows the interest of the method and its robustness.
Critical weaknesses
  • If the methodology part explains why both temporal and structural aspect are taking into account, it could be nice to have qualitatives results like plot of windows considering close or not to see what are the differences. It could help to see concretely what other methods are missing and what Tex Mex is bringing.

  • In the same direction the ablation study didn't really investigate the interest of structural information or temporal information. It could be nice to see when one or the other is important or not.

  • The claim of contrastive learning for time series is strong since the experiments are only done on "medical" time series. These time series are very typical, I'm wondering if the conclusions stand for other type of time series.

Questions for authors
  • On the first table, Pulse arrives very close to the performance of Tex-Mex. On the third table, Pulse-PPG gives way less good performance than Tex-Mex. I'm wondering what are the methodological difference here ? Since in the last experiment you are not comparing the same backbone for all the method, do you think the backbones plays a important role and you make the best choice for Tex-Mex one ?
4 comments
AuthorsOpenReview ↗

Response to Reviewer tesk (1/2)

We thank the reviewer for their thoughtful feedback and for recognizing TEX-MEX's motivation and empirical strengths. In response, we have added qualitative examples and clarified several aspects of our evaluation. In brief:

  • Qualitative examples. We added nearest-neighbor visualizations comparing TEX-MEX, DTW, and REBAR, complementing our quantitative similarity analysis (§1).
  • Morphology-derived similarity and temporal context. We clarified how our ablations and baseline comparisons examine the contributions of both, and added new similarity measure ablations (§2).
  • Generalizability. We highlighted the diversity of our signal domains, including non-medical activity recognition, and refined the scope of our conclusions in the paper (§3).
  • Foundation model comparisons. We clarified why the task-specific PULSE benchmark and broader foundation model evaluation reveal different performance margins, and addressed the role of the backbone (§4).

We address each point below.

1. Qualitative examples

"It could be nice to have qualitatives results like plot of windows considering close or not to see what are the differences. It could help to see concretely what other methods are missing and what Tex Mex is bringing."

We agree that qualitative examples can help make the similarities identified by TEX-MEX more concrete. We have added an appendix figure comparing nearest-neighbor retrieval by TEX-MEX, DTW, and REBAR on held-out ECG and PPG windows (Appendix C, Figure 4). For each query, we show the top-ranked windows under each measure alongside randomly sampled windows.

We chose ECG and PPG as, among the four domains in our PULSE evaluation, differences between their retrieved windows are the most easily interpretable visually. In ECG, atrial fibrillation shows up as irregular beat-to-beat timing, which the figure plots below each trace. In PPG, the example contrasts stress with other affective states in WESAD. These examples provide visual context for the windows each measure retrieves, while our quantitative class-membership analysis in Fig. 3 provides a more systematic assessment.

2. Morphology-derived similarity and temporal context

"In the same direction the ablation study didn't really investigate the interest of structural information or temporal information. It could be nice to see when one or the other is important or not."

Our experiments address both, and we have revised the text to identify the relevant ablations and comparisons more explicitly (Section 4.2, lines 400–420).

To examine the contribution of morphology-derived similarity, our "w/o graded weighting" ablation retains temporal shifts but removes similarity-based grading of the positive, negative, and frame-separation terms. This degrades 11 of 12 metrics, by up to 17.1 AUROC (Table 3; also shown in the General Response). Conversely, fixing δ = 0 removes temporal shifting entirely and is the most damaging variant, though it lacks an external distance measure. A cleaner view of the value of temporal context comes from comparisons against morphology-based methods such as REBAR, which pair similar windows using an external distance but no temporal context, and which TEX-MEX broadly outperforms in our benchmarks.

Together, these results support the importance of composing morphology-derived similarity with temporal context. We have made the correspondence between these experiments and the two components more explicit in the revised paper (Section 4.2, lines 400–407).

(continued in part 2/2)

AuthorsOpenReview ↗

Response to Reviewer tesk (2/2)

3. Generalization across time series domains

"The claim of contrastive learning for time series is strong since the experiments are only done on 'medical' time series. These time series are very typical, I'm wondering if the conclusions stand for other type of time series."

We agree with the reviewer that our conclusions are best stated within physiological and wearable sensing, and we have accordingly focused the paper's claims to biosignals, particularly mHealth and wearable data (Section 1, lines 34–43; Section 6, lines 521–526). At the same time, our evaluation spans five substantially different signal domains: PPG, ECG, EEG, CGM, and human activity recognition (HAR). These differ considerably in signal morphology, temporal dynamics, sampling characteristics, and downstream tasks. HAR, for example, is a non-medical time series task using accelerometry for everyday activity recognition. This diversity suggests promise for other kinds of time series, and evaluating TEX-MEX in those settings remains future work. We note this in the paper's limitations (Section 6, lines 524–525).

4. PULSE and foundation model comparisons

"On the first table, Pulse arrives very close to the performance of Tex-Mex. On the third table, Pulse-PPG gives way less good performance than Tex-Mex. I'm wondering what are the methodological difference here ? Since in the last experiment you are not comparing the same backbone for all the method, do you think the backbones plays a important role and you make the best choice for Tex-Mex one ?"

We appreciate the reviewer raising this question, as the two comparisons serve different purposes. First, PULSE (Chen et al., 2026) and Pulse-PPG (Saha et al., 2025) are distinct approaches, which we now clarify in the paper (Section 4, footnote, line 269). Table 1 compares PULSE alongside other SSL methods using a shared encoder and evaluation protocol across several signal domains. Its PPG evaluation, however, is limited to WESAD, with hyperparameters selected for that task using validation performance. While this provides evidence of TEX-MEX's effectiveness across diverse domains, the PPG result reflects performance on a single task under task-specific tuning.

Table 4 instead compares pretrained foundation models, including Pulse-PPG, across 11 downstream tasks and multiple datasets, evaluating how broadly their representations transfer. TEX-MEX's advantage over Pulse-PPG on WESAD is relatively small compared with the substantially larger gains on other tasks, such as hypertension prediction. The broader evaluation therefore reveals advantages in general-purpose representations that are less apparent from PULSE's WESAD evaluation alone.

Regarding the backbone, our PPG foundation model uses a standard 1D ResNet and a typical pretraining setup, without relying on a larger or specialized encoder. Table 4 compares complete foundation model systems under their respective configurations and tuned hyperparameters, while the controlled SSL comparisons in Tables 1-2 separately establish the effectiveness of the TEX-MEX objective under shared encoder settings.

References

  • Chen, Y., Xu, M. A., Rehg, J. M., and Rozell, C. J. Self-supervised dynamical system representations for physiological time-series. ICML, 2026.
  • Saha, M., Xu, M. A., Mao, W., Neupane, S., Rehg, J. M., and Kumar, S. Pulse-PPG: An open-source field-trained PPG foundation model for wearable applications across lab and field settings. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 9(3):1–35, 2025.
  • Schmidt, P., Reiss, A., Duerichen, R., Marberger, C., and Van Laerhoven, K. Introducing WESAD, a multimodal dataset for wearable stress and affect detection. Proceedings of the 20th ACM International Conference on Multimodal Interaction, 400–408, 2018.
  • Xu, M., Moreno, A., Wei, H., Marlin, B. M., and Rehg, J. REBAR: Retrieval-based reconstruction for time-series contrastive learning. ICLR, 2024.

I thank the authors for their answers. The paper is good and even stronger with the new experiments and visualisations and it confirms my score of 3. If the good results encourage me to give a weak accept, the study remains simple (but efficient) and a score of 3 describe well the work done.

We sincerely thank the reviewer for their time, for engaging with our response, and for their positive assessment. We appreciate the suggestions on qualitative examples and the ablation analysis, which have strengthened the paper.

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

This paper proposes TEX-MEX, a similarity-guided temporal contrastive learning framework for time-series representation learning. The method aims to unify structural and temporal similarity by using encoder-derived representations to estimate relationships between windows and constructing contrastive objectives over their temporally shifted counterparts.

The authors evaluate TEX-MEX across five time-series domains, including PPG, ECG, EEG, accelerometry, and continuous glucose monitoring. The experimental results demonstrate competitive or state-of-the-art performance on a broad range of downstream tasks. Furthermore, the authors train a lightweight PPG foundation model with only 560K parameters, achieving strong performance compared with substantially larger models.

Critical strengths

The paper provides extensive experimental evaluation across different signal domains and downstream tasks. The comparisons with contrastive, reconstructive, and other self-supervised methods demonstrate the effectiveness and general applicability of the proposed objective. The ablation studies also provide useful evidence for the contributions of graded weighting, temporal shifting, and temporal frame separation.

Critical weaknesses

The claim doesn't hold that contrastive learning in TS mainly contains structural and temporal similarities. Since 2020, there are hundreds of contrastive learning algorithms in time series SSL. Each augmentation could provide one similarity in the positive-pair definition: jittering, flipping, scaling, masking etc. It's good to combine structural and temporal, but we need to make it clear that the literature in this domain is much broader.

Questions for authors
  1. Could the authors further clarify the specific novelty of TEX-MEX relative to existing methods that already incorporate both structural and temporal similarity, particularly SoftCLT?

  2. Can the authors provide additional evidence that the learned representations capture the temporal-context relationships described in the Introduction, beyond improved class-based nearest-neighbor similarity?

  3. How would TEX-MEX perform if the encoder-derived similarity weights were replaced with fixed similarity measures, such as REBAR or DTW, while keeping the remaining objective unchanged?

  4. Could the authors clarify how much of the lightweight PPG foundation model's performance improvement is attributable to the proposed learning objective rather than the modified architecture and pretraining configuration?

2 comments
AuthorsOpenReview ↗

Response to Reviewer Hnru (1/2)

We thank the reviewer for their constructive feedback and thoughtful questions. We have addressed their concerns through additional experiments and revisions to the paper. In brief:

  • Novelty relative to SoftCLT. TEX-MEX unifies morphology-derived similarity and temporal context within the same contrastive pair, allowing similar windows across recordings to contribute their varied temporal contexts (§2).
  • Encoder-based similarity. Following the reviewer's suggestion, we evaluated DTW and REBAR as alternative graders. Both reduce performance and substantially increase training time (§4).
  • Augmentation-based methods. We have clarified the scope of our motivation and expanded the Introduction and Related Work to discuss augmentation-based approaches (§1).
  • Foundation model attribution. Our PPG foundation model uses a standard 1D ResNet and a conventional pretraining setup. Controlled SSL comparisons across five signal domains independently establish the objective's effectiveness (§5).

We address each question below.

1. Augmentation-based methods

"The claim doesn't hold that contrastive learning in TS mainly contains structural and temporal similarities… Each augmentation could provide one similarity in the positive-pair definition: jittering, flipping, scaling, masking etc."

We agree that augmentation-based methods form a large and important family of approaches, and have clarified that our discussion of temporal and morphological similarity specifically concerns methods that define positive pairs directly from the data rather than through augmentations. While augmentations provide useful supervision, they can also distort physiologically meaningful features or alter time series semantics (Xu et al., 2024). Like prior works, TEX-MEX avoids this risk by constructing its pairs from observed windows without relying on augmentations. We have revised the Introduction and Related Work to clarify this distinction and discuss augmentation-based approaches explicitly (Section 1, lines 71–72; Section 2, lines 118–122).

2. Novelty relative to SoftCLT

"Could the authors further clarify the specific novelty of TEX-MEX relative to existing methods that already incorporate both structural and temporal similarity, particularly SoftCLT?"

Unlike SoftCLT's separate instance-wise and temporal losses, TEX-MEX pairs each anchor window with the temporal context of morphologically similar windows across recordings, capturing relationships that are only identified when morphological and temporal relations are considered simultaneously. SoftCLT's instance-wise loss compares different recordings at the same timestamp, while its temporal loss compares timestamps within a single input (Lee et al., 2024). These losses therefore capture cross-recording similarity and temporal proximity separately. TEX-MEX instead composes these relationships within the same contrastive pair at the window level, allowing temporal context from other recordings to contribute according to the similarity of the source windows. Empirically, TEX-MEX outperforms SoftCLT on 18 of 20 results, by up to 14.1 AUROC (Tables 1-2). Moreover, removing similarity-based grading degrades 11 of 12 metrics, by up to 17.1 AUROC (Table 3; also shown in the General Response). We have revised the paper's discussion to illustrate this distinction more directly (Section 1, lines 80–90; Section 2, lines 129–137), with further explanation in the General Response ("What distinguishes TEX-MEX?").

TEX-MEX's learned similarity measure is also more effective and computationally efficient. SoftCLT uses fixed DTW distances, which can be precomputed across all pairs of training inputs at a cost that scales quadratically with dataset size, or computed during training at additional cost (Lee et al., 2024). In our ablations, computing DTW within each batch increases per-step training time by 5.9x and degrades 11 of 12 metrics, by up to 7.8 AUROC (revised Table 3). TEX-MEX instead uses encoder-derived similarity, reusing representations already computed in the forward pass. SoftCLT also reports that embedding-based contrastive approaches such as NNCLR and ASCL degrade TS2Vec's performance (Sec. 4.6), whereas encoder-derived similarity works effectively within TEX-MEX's different pair construction.

(continued in part 2/2)

AuthorsOpenReview ↗

Response to Reviewer Hnru (2/2)

3. Temporal-context representation evidence

"Can the authors provide additional evidence that the learned representations capture the temporal-context relationships described in the Introduction, beyond improved class-based nearest-neighbor similarity?"

Our ablations provide evidence that these relationships contribute to downstream performance. The "w/o graded weighting" variant preserves the same temporal shifts but makes the anchor's own shift its only positive, removing positive supervision from the temporal contexts of similar windows across recordings. This degrades 11 of 12 metrics, by up to 17.1 AUROC (Table 3).

A convincing direct evaluation would ideally use dense, fine-grained annotations identifying meaningful short-term state transitions and temporal-context relationships, which our datasets unfortunately do not provide. Evaluating these relationships solely through the encoder's learned similarity would not provide an independent measure of their semantic relevance. We have therefore revised the Introduction to clarify that the exercise and recovery example illustrates the kind of relationship constructed by the objective, rather than a property directly verified in the learned representations (Section 1, lines 85–90).

4. Encoder-based similarity versus DTW/REBAR

"How would TEX-MEX perform if the encoder-derived similarity weights were replaced with fixed similarity measures, such as REBAR or DTW, while keeping the remaining objective unchanged?"

We thank the reviewer for suggesting this experiment, which has strengthened our evaluation of the encoder-based similarity measure. The encoder-based grader outperforms both fixed similarity measures while being substantially more efficient. We replaced the encoder-based grader with DTW and REBAR while keeping the remaining objective unchanged. The DTW and REBAR graders degrade 11 and 10 of 12 metrics, by up to 7.8 and 7.7 AUROC, while per-step training time increases by 5.9x and 13x, respectively (revised Table 3).

Because these distance measures have different numerical scales, we evaluated seven values of the grading temperature τw\tau_w, which controls how distances are converted into the contrastive weights w+w^+ and w−w^- in the TEX-MEX loss. We report the best observed test AUROC across these values, giving the fixed graders the benefit of post hoc selection, while TEX-MEX retains its default τw\tau_w. Despite this favorable comparison, both fixed graders underperform TEX-MEX.

5. PPG foundation model attribution

"Could the authors clarify how much of the lightweight PPG foundation model's performance improvement is attributable to the proposed learning objective rather than the modified architecture and pretraining configuration?"

Our PPG foundation model uses a standard 1D ResNet architecture and a typical pretraining setup, without relying on a specialized encoder. We compare against foundation models that likewise use large-scale pretraining and configurations tailored to their respective approaches, including PPG-specific methods and models pretrained on multimodal datasets. Pulse-PPG is a particularly direct comparison, using the same MOODS pretraining dataset and splits with a similar ResNet architecture. Despite this conventional setup, TEX-MEX achieves stronger performance across most downstream tasks (Table 4).

We evaluate the objective's contribution separately in Tables 1-2, comparing TEX-MEX against other SSL methods under controlled encoder architectures and evaluation protocols across five signal domains. TEX-MEX leads in both evaluations, supporting the effectiveness of the objective independently of our PPG foundation model setup.

References

  • Dwibedi, D., Aytar, Y., Tompson, J., Sermanet, P., and Zisserman, A. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. ICCV, 2021.
  • Lee, S., Park, T., and Lee, K. Soft contrastive learning for time series. ICLR, 2024.
  • Neupane, S., Saha, M., Ali, N., Hnat, T., Samiei, S. A., Nandugudi, A., Almeida, D. M., and Kumar, S. Momentary stressor logging and reflective visualizations: Implications for stress management with wearables. CHI, 2024.
  • Saha, M., Xu, M. A., Mao, W., Neupane, S., Rehg, J. M., and Kumar, S. Pulse-PPG: An open-source field-trained PPG foundation model for wearable applications across lab and field settings. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 9(3):1–35, 2025.
  • Xu, M., Moreno, A., Wei, H., Marlin, B. M., and Rehg, J. REBAR: Retrieval-based reconstruction for time-series contrastive learning. ICLR, 2024.
  • Yue, Z., Wang, Y., Duan, J., Yang, T., Huang, C., Tong, Y., and Xu, B. TS2Vec: Towards universal representation of time series. AAAI, 2022.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to Reviewers (1/2)

We thank all reviewers for their detailed feedback. Reviewers found the motivation clear and the results strong and robust across modalities (tesk), noted the extensive evaluation across signal domains and downstream tasks as demonstrating the objective's effectiveness and general applicability (Hnru), and agreed that the motivation of combining semantic and temporal information while avoiding augmentations that risk distorting time series is desirable (U2kw). Below, we clarify the core distinction in TEX-MEX's approach, which ultimately enables its simplicity, performance, and computational efficiency. We then summarize our new experiments and revisions to the paper. Revisions in the updated PDF are highlighted in blue.

What distinguishes TEX-MEX?

TEX-MEX pairs each anchor window with the temporal context of morphologically similar windows across recordings. As a result, it can capture relationships that are only identified when morphological and temporal relations are considered simultaneously, which, to our knowledge, no prior method does.

We describe TEX-MEX relative to prior contrastive approaches that select positives from the dataset:

  1. Temporal approaches (e.g., CPC, TNC) relate a window to its temporal context within a recording. This context can include morphologically dissimilar windows, such as the recovery that follows exercise, but it only captures temporal context within a single recording. Positive pairs should capture the range of variation among semantically related windows (Xu et al., 2024), and much of that variation lies across recordings (i.e., individuals): in wearable accelerometry, for example, the similarity between two instances of the same activity that are present in two different recordings cannot be captured by local temporal relationships alone.
  2. Morphological approaches (e.g., REBAR, RelCon) provide the ability to make comparisons across recordings (i.e., individuals) by pairing morphologically similar windows regardless of their distance in time, within and across recordings. In doing so, however, they miss capturing the relationship between windows that are related by time yet morphologically dissimilar, such as a post-meal glucose rise and the decline that follows it in CGM.
  3. SoftCLT is the closest prior work: like TEX-MEX, it incorporates both morphological and temporal similarity and weights pairs softly. It does so by softening TS2Vec's two existing losses, each of which relates only one kind of pair: different recordings at the same timestamp, weighted by DTW, or nearby timestamps within a single input window, weighted by their distance in time. Every SoftCLT pair therefore spans either across recordings or temporal shifts within a recording, never both simultaneously. SoftCLT thus inherits the limitation of temporal approaches: temporal relationships are only considered within a recording, missing the variation in context surrounding similar patterns across recordings.
  4. TEX-MEX overcomes this limitation by unifying morphological similarity and temporal context within the same pair: the encoder's similarity between the anchor and other sampled windows, across recordings, determines whose temporal context enters the contrast and with what weight. Similarity is judged on these source windows, while their time-shifted counterparts become the positives, so a positive need not be morphologically similar to the anchor. As an illustrative example, a PPG window recorded during exercise could be paired not only with the recovery that follows it in its own recording, but with the recoveries that follow similar exercise in other recordings, even though recovery is morphologically distinct from exercise. Since recovery varies widely across exercise sessions and across people, the objective thereby draws on many instances of recovery rather than the single one in the anchor's own recording. As a soft contrastive objective, TEX-MEX weights each pair by how similar its source window is to the anchor, e.g., for a sprint window, the recovery after another sprint counts more than the recovery after a light jog.

These design choices are borne out empirically: across five signal domains, TEX-MEX broadly achieves state-of-the-art performance against contrastive, reconstructive, JEPA-based, and modality-specific methods, including SoftCLT on 18 of 20 results, by up to 14.1 AUROC (Tables 1-2, 4).

We thank the reviewers for flagging where TEX-MEX's distinction with prior work was unclear, and have revised the Introduction and Related Work accordingly (Section 1, lines 71–90; Section 2, lines 118–137).

(continued in part 2/2)

AuthorsOpenReview ↗

General Response to Reviewers (2/2)

Scope

TEX-MEX targets biosignals, which are typically recorded as long, continuous recordings from individuals. Learning effective representations of such signals requires comparing windows both within one individual's recording and across recordings from different individuals, because individual differences play an important role in health-related tasks. TEX-MEX is primarily designed for this setting, particularly mHealth and wearable data, and we have scoped our claims accordingly (Section 1, lines 34–43; Section 6, lines 521–526).

Updated Experiments

We expanded our ablation study (updated Table 3) and added qualitative examples. In the table below, A and P denote AUROC and PRAUC on CGM metabolic subphenotyping, for β-cell dysfunction (β) and insulin resistance (IR) in the Home, Venous (Ven.), and Venous→Home transfer (V→H) settings.

VariantA: Home βA: Home IRA: Ven. βA: Ven. IRA: V→H βA: V→H IRP: Home βP: Home IRP: Ven. βP: Ven. IRP: V→H βP: V→H IRTime/step
TEX-MEX.980.952.855.796.978.952.972.946.882.848.967.9436.69 ms
Temporal only (w/o grading)−.042−.097−.049+.001−.171−.142−.046−.090−.118−.008−.159−.125−0.9%
[NEW] w/o graded weighting in N−.018−.066+.019+.004−.016−.066−.020−.044−.009−.015−.015−.041−0.0%
w/o frame separation H−.018−.066+.018+.004−.012−.057−.020−.044−.010−.015−.010−.033−0.2%
w/o stop-gradient−.037−.079+.019−.010−.025−.069−.034−.052−.009−.019−.025−.047+2.3%
δ = 0−.227−.198−.398−.326−.220−.146−.265−.203−.492−.404−.232−.138−0.2%
EMA grader (m = 0.999)−.034−.090+.009+.016−.032−.090−.031−.057−.020−.009−.026−.054+35.7%
[NEW] DTW grader†−.018−.066−.006+.001−.012−.078−.020−.044−.030−.016−.014−.063+487%
[NEW] REBAR grader†−.034−.077+.025+.004−.012−.060−.031−.051−.010−.015−.009−.035+1229%

†Grading temperature selected as the best of seven values by test AUROC, which favors these baselines; TEX-MEX uses its default. The top row reports absolute performance; all other rows report the absolute change from TEX-MEX (revised Table 3). Throughout our responses, differences are stated as absolute percentage points, e.g., −.171 is written as a drop of 17.1 AUROC.

  1. Fixed similarity measures in place of the encoder (Hnru). Replacing the encoder's self-similarity with DTW or REBAR degrades 11 and 10 of 12 metrics, by up to 7.8 and 7.7 AUROC, respectively, at much higher cost (item 3).
  2. Component ablations (U2kw, Hnru). Our new uniform-negative-weight ablation degrades 10 of 12 metrics. In the existing ablations, removing graded weighting costs up to 17.1 AUROC, under venous→home transfer; removing H or the stop-gradient costs up to 6.6 and 7.9 AUROC; and an EMA grader costs up to 9.0 AUROC.
  3. Computational cost (U2kw). We measured per-step training time for every ablation (updated Table 3). TEX-MEX's own loss terms add no measurable cost: making N's weights uniform, removing H, or removing graded weighting changes per-step time by less than 1%. The cost lies in the alternatives: an EMA grader adds 36%, while grading with DTW or REBAR multiplies per-step time by 5.9x and 13x, respectively. All variants were timed under identical conditions on a single H100 GPU. DTW is computed exactly, with warping limited to 25% of the window length. REBAR uses a pretrained distance model, whose training time is excluded from the per-step cost.
  4. Qualitative examples (tesk). A new appendix figure (Appendix C, Figure 4) shows, for held-out ECG and PPG queries, illustrative nearest neighbors retrieved by DTW, REBAR, and TEX-MEX, alongside randomly sampled windows.

References

  • Lee, S., Park, T., and Lee, K. Soft contrastive learning for time series. ICLR, 2024.
  • Oord, A. van den, Li, Y., and Vinyals, O. Representation learning with contrastive predictive coding. arXiv:1807.03748, 2018.
  • Tonekaboni, S., Eytan, D., and Goldenberg, A. Unsupervised representation learning for time series with temporal neighborhood coding. arXiv:2106.00750, 2021.
  • Xu, M., Moreno, A., Wei, H., Marlin, B. M., and Rehg, J. REBAR: Retrieval-based reconstruction for time-series contrastive learning. ICLR, 2024.
  • Xu, M., Narain, J., Darnell, G., Hallgrimsson, H., Jeong, H., Forde, D., Fineman, R., Raghuram, K., Rehg, J., and Ren, S. RelCon: Relative contrastive learning for a motion foundation model for wearable data. ICLR, 2025.
  • Yue, Z., Wang, Y., Duan, J., Yang, T., Huang, C., Tong, Y., and Xu, B. TS2Vec: Towards universal representation of time series. AAAI, 2022.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.