acceptodds
Under review as a conference paper at ICLR 2027

META: Meta-learning Enhanced Temporal-Frequency Alignment for Time Series Domain Adaptation

Abstract

Time series classification (TSC) underpins clinical decision support, yet differences in patient demographics and institutional data collection workflows induce domain shift that severely degrades model performance on the target domain. Unsupervised domain adaptation (UDA) mitigates this by aligning distributions, but we show that the two dominant alignment paradigms each leave a distinct structural blind spot. First, joint maximum mean discrepancy (JMMD) constrains only class-conditional means, leaving mean-preserving drifts in its null space unpenalized, which drifts enlarge intra-class dispersion, directly reducing discriminability. Second, contrastive learning pulls all same-class samples toward a single centroid, and this non-uniform attraction radially stretches each sub-mode, enlarging sub-modal dispersion and destroying the multi-modal structure that carries clinically meaningful heterogeneity. We address both blind spots with META, a unified time-series UDA framework comprising two modules: (i) Time-Frequency Contrastive Learning leverages the temporal dependency of time and the stability of frequency to learn complementary contextual representations. We prove that at its global optimum same-class features collapse to class centers forming a regular simplex. (ii) MetaMMD treats the Gaussian kernel bandwidths of MMD as learnable parameters optimized via bi-level optimization against a domain discrimination loss, which differentiates bandwidths so that same-sub-mode cross-domain kernel weights remain bounded below by a positive constant while cross-sub-mode weights vanish. We prove that this shrinks within-sub-mode variance exponentially and drives sub-modal dispersion toward zero without merging sub-modes. Experiments on eight datasets across four tasks demonstrate that META consistently outperforms state-of-the-art UDA and time-series UDA baselines.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Time Series Classification, Unsupervised Domain Adaptation, Contrastive Learning, Meta-Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_gYt5OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose to use meta-learning with a domain discriminator to alleviate the issues with JMMD and contrastive learning in having multi-model features for multiple domains. They provide a theoretical framework to show these issues as well in theory and how their method does not suffer from the same issues. On multiple datasets they outperform the state-of-the-art

Critical strengths

The paper is well motivated, solves an important problem and is thoroughly evaluated on multiple datasets. The authors also include a theoretical analysis and ablations.

Critical weaknesses

Measurement of Sub-Modal dispersion: They measure this parameter using a GMM distribution assumption. Now their model assumption is also that they following some kind of anisotropic gaussians with sigma_ij as learnt bandwiths. This naturally makes their method perform better than competing methods without this assumption. There is no other metric be this quantitative like OT-based or qualitative like t-sne to gather some more insight.

The inner loop updates are computed using a finite difference style formulation, however it is unclear how close this is approximation is to the real gradient even in praxis. There is no theory behind it and no error estimate. There is not even a numerical experiment showing that during trainig the error is in the ballpark of x% of the norm of the gradient.

Questions for authors

Can you also visualize confirm that the modes are actually separated and do you have some insight into which features they found as different modes?

Can you elaborate on the finite style derived gradient computation and how close they are to the real gradient?

Minor: Check the citations, the core competitor JMMD has the wrong year.

5 comments
AuthorsOpenReview ↗

Response to Reviewer gYt5 W1

We sincerely thank the reviewer for the positive assessment of our motivation, theoretical analysis, and experimental evaluation, and for the insightful suggestions on assessing sub-modal structure. Your comments helped us strengthen the evaluation beyond within-sub-mode compactness. We have uploaded the revised manuscript, with all changes highlighted in blue. Below we provide our detailed response.

W1. Measurement of Sub-Modal dispersion: They measure this parameter using a GMM distribution assumption. Now their model assumption is also that they following some kind of anisotropic gaussians with sigma_ij as learnt bandwiths. This naturally makes their method perform better than competing methods without this assumption. There is no other metric be this quantitative like OT-based or qualitative like t-sne to gather some more insight.

Response. Thank you for raising this concern. We clarify the distinction between the Gaussian kernel used for training and the GMM used for evaluation, and provide additional density-based clustering, geometric measurements, and visualizations.

1. A Gaussian kernel is not a Gaussian-mixture density assumption.

Since σij\sigma_{ij} is an MMD interaction scale rather than a fitted GMM covariance, and MetaMMD never optimizes a GMM likelihood, the Gaussian kernel confers no built-in advantage under GMM evaluation, though we agree an independent evaluation is indeed valuable. Our revised evaluation separates fixed-reference geometry from independent cluster discovery:

  • Fixed-reference geometry. We construct fixed reference groups across all methods by applying PCA to standardized time–frequency descriptors and fitting a classwise GMM (BIC-selected K∈{1,2,3,4,5}K\in\{1,2,3,4,5\}). We then compute empirical covariance on ℓ2\ell_2-normalized embeddings within these frozen groups.

  • Independent cluster discovery. We evaluate classwise GMM/BIC clustering and HDBSCAN on the same normalized embeddings, using identical settings across methods. HDBSCAN uses minsamples=5{min_samples}=5 and minclustersize=max⁡(10,⌈0.02nc⌉){min_cluster_size}=\max(10,\lceil0.02n_c\rceil).

Without GMM, HDBSCAN independently detects multiple populated clusters. The following examples show that the observed multiplicity of clusters is not solely an artifact of fitting Gaussian components.

HHAR_SA 1→61\to6, cycling/class 0: 483 samples; Appendix G.2, Fig. 8.

MethodHDBSCAN clustersNon-noise cluster sizesNoise samplesNoise rate
w/o MetaMMD2416, 2938 / 4837.9%
BSB-MMD2135, 13335 / 48369.4%
META478, 13, 241, 14137 / 48328.4%

Without fitting a Gaussian mixture to discover these clusters, HDBSCAN identifies four clusters containing 346/483 samples under META.

WISDM 4→154\to15, class 5: 107 samples; additional classwise heatmap analysis.

MethodHDBSCAN clustersNon-noise cluster sizesNoise samplesNoise rate
w/o MetaMMD239, 2840 / 10737.4%
META246, 529 / 1078.4%

HDBSCAN identifies two populated clusters containing 98/107 samples under META. The two reference groups have distinct dominant destinations, with 46/53 and 52/54 samples entering their respective clusters.

HAR 12→1612\to16, class 5: 91 samples; Fig. 6.

MethodHDBSCAN clustersNon-noise cluster sizesNoise samplesNoise rate
w/o MetaMMD323, 42, 251 / 911.1%
META426, 20, 16, 1811 / 9112.1%

HDBSCAN independently detects four clusters containing 80/91 samples under META. In particular, all 18 R1 samples and 16 R4 samples enter one cluster without MetaMMD but occupy distinct clusters with META.

AuthorsOpenReview ↗

Response to Reviewer gYt5 W1

3. Additional metrics distinguish compactness from separation.

We appreciate the suggestion of OT-based evaluation. We prioritized the added metrics because they directly test the failure mode at issue: whether distinct same-class groups remain separated. So we introduced additional metrics in Appendix E.4: Dmin⁡D_{\min}, the minimum distance between two same-class sub-mode centers; RsubR_{\rm sub}, that distance relative to within-sub-mode spread; Silhouette, within-group cohesion versus separation; Submode Match, agreement with the nearest same-class source sub-mode center; and Merge Rate, the fraction of same-class reference pairs assigned to the same dominant non-noise cluster.

Fig. 4 reports averages over ten transfer scenarios per dataset. Each entry below is Vsub↓V_{\rm sub}\downarrow / Rsub↑R_{\rm sub}\uparrow.

DatasetSourceJANw/o MetaMMDMETA
HHAR_SA0.0257 / 0.31040.0563 / 0.27820.0296 / 0.32220.0261 / 0.3439
WISDM0.0331 / 0.53200.0506 / 0.26310.0314 / 0.30620.0184 / 0.3527
HAR0.0398 / 0.16680.0568 / 0.26090.0288 / 0.37840.0239 / 0.3604
Sleep-EDF0.0602 / 0.48250.0265 / 0.43760.0210 / 0.56430.0049 / 0.5886

Table 4 provides a task where compactness, absolute separation, and correspondence improve together:

EMG 1→31\to3Vin↓V_{\rm in}\downarrowVsub↓V_{\rm sub}\downarrowDmin⁡↑D_{\min}\uparrowRsub↑R_{\rm sub}\uparrowSilhouette ↑Submode Match ↑Merge Rate ↓
w/o MetaMMD0.12890.08400.00830.0256−0.028943.43%56.60%
META0.09070.07270.04780.15430.038658.49%18.87%

Here, the groups become tighter while even their closest centers move farther apart. Improved Silhouette, Submode Match, and Merge Rate provide additional evidence that the lower dispersion is accompanied by retained distinction between groups.

4. Training dynamics and loss-balance analysis provide complementary evidence (Fig. 5b–c).

For WISDM 18→2018\to20, Fig. 5b tracks DinterD_{\rm inter}, the minimum same-class sub-mode center distance. Selected values are shown below, averaged over three seeds; discriminator pretraining ends at epoch 20.

Dinter↑D_{\rm inter}\uparrowEpoch 0Epoch 20Epoch 24Epoch 50Epoch 70
w/o MetaMMD0.02900.01550.00340.01050.0174
META0.02900.01550.00400.01480.0276

Both methods initially contract, but META subsequently recovers greater separation, reaching 0.0276 versus 0.0174 at epoch 70. This supports improved separation during later joint optimization, rather than an absence of transient contraction.

On WISDM 2→322\to32, Fig. 5c varies ρ=λSCon/λTCon\rho=\lambda_{\rm SCon}/\lambda_{\rm TCon} while keeping the sum of these weights fixed. The table reports all five configurations for the plotted seed.

ρ\rhow/o MetaMMD: Vsub↓V_{\rm sub}\downarroww/o MetaMMD: Dinter↑D_{\rm inter}\uparrowMETA: Vsub↓V_{\rm sub}\downarrowMETA: Dinter↑D_{\rm inter}\uparrow
00.03130.00820.04070.0092
0.250.04440.00780.03450.0221
0.50.03860.01520.03760.0258
10.04390.00930.04060.0105
∞\infty (w/o TCon)0.03840.00700.03480.0103

META achieves both lower dispersion and greater separation at four of the five matched settings. At ρ=0\rho=0, separation improves but dispersion increases. These results show why compactness and separation should be evaluated jointly.

5. Assignment-based analyses assess merging across transfer scenarios.

Appendix G.3, Fig. 9 reports pooled, all-class Merge Rate (%):

WISDM26→221→317→3020→302→3212→725→2912→1918→204→15Mean
w/o MetaMMD18.1815.0013.8912.5010.2610.007.504.764.760.009.69
META6.8215.002.784.1710.260.000.004.7614.290.005.81
EMG1→31→20→30→20→13→12→32→12→0Mean
w/o MetaMMD56.6052.9442.1125.0021.6216.228.168.117.3226.45
META18.8723.532.6319.6410.8116.228.1616.2221.9515.34

Mean Merge Rate decreases from 9.69% to 5.81% on WISDM and from 26.45% to 15.34% on EMG. On WISDM, average noise changes only from 33.65% to 33.82%. Improvements are not universal, as the individual scenarios show.

The Sankey diagrams (Fig. 6) and heatmaps (Appendix G.2, Fig. 8) trace individual reference groups: merged pairs decrease from 3/10 to 0/10 on HAR 12→1612\to16, from 1/3 to 0/3 on WISDM 18→2018\to20 (both class 5), and from 10/10 to 3/10 on HHAR_SA 1→61\to6 (cycling), with noise reported separately. The class-specific improvement on WISDM 18→2018\to20 coexists with an increased all-class merge rate, reflecting different outcomes across classes.

AuthorsOpenReview ↗

Response to Reviewer gYt5 W2

W2. The inner loop updates are computed using a finite difference style formulation, however it is unclear how close this is approximation is to the real gradient even in praxis. There is no theory behind it and no error estimate. There is not even a numerical experiment showing that during trainig the error is in the ballpark of x% of the norm of the gradient.

Response. We thank the reviewer for requesting a quantitative assessment of the approximation. We have added gradient diagnostics in Appendix G.5, Eqs. (57)–(59), and Fig. 11. We clarify that finite differences approximate the mixed derivative used in the outer bandwidth update and the inner model updates use backpropagation and Adam.

1. Why an approximate update can be useful, and what finite differences approximate.

Optimization does not require every update to reproduce the exact gradient in both direction and magnitude. For example, for a nonzero gradient gg, the estimate g~=cg\widetilde g=cg, with c>0c>0, has the same direction and changes the effective step size, although its relative error is ∣c−1∣|c-1|. Nevertheless, this observation does not establish the quality of our estimator: its direction and magnitude must be measured against a specified reference.

Our approximation uses a centered finite difference to estimate a local mixed-derivative–vector product, following the computational principle used in DARTS (Liu et al., ICLR 2019). To state its error precisely, fix the snapshot, the support/query batches, and the outer-gradient direction vv. Let b=gϕ(θ)b=g_\phi(\theta) denote the predicted support bandwidths, and hold their input embeddings fixed during the two perturbations, as in the implementation. With Jϕ=∂b/∂ϕJ_\phi=\partial b/\partial\phi, define

F(t)=Jϕ⊤∇bLb(θ+tv,b).F(t)=J_\phi^\top\nabla_b L_b(\theta+tv,b).

The local automatic-differentiation reference and its finite-difference approximation are

\widetilde g_\epsilon=-\eta\frac{F(\epsilon)-F(-\epsilon)}{2\epsilon}.$$ If $F$ is three times continuously differentiable on the perturbation interval and $\sup_{|t|\leq\epsilon}\|F'''(t)\|\leq M$, Taylor's theorem gives $$\|\widetilde g_\epsilon-g_{\rm local}\| \leq\frac{\eta M}{6}\epsilon^2.$$ This is a conditional local truncation-error bound, rather than a global smoothness guarantee for a network with piecewise-smooth operations. Finite precision also matters: if each evaluation of $F$ has error at most $u$, the additional error is bounded by $\eta u/\epsilon$. Thus, making $\epsilon$ arbitrarily small need not improve numerical accuracy. **2. Experimental verification of the finite difference supergradient** In **Fig. 11**, on WISDM $2\to32$, gradients are measured every five epochs on fixed probe batches with $\epsilon=0.001$. We compare $g_{\rm local}$ (local mixed derivative) with $\widehat{g}_K$ (differentiating through $K$ inner Adam updates from the snapshot's optimizer state), a reference for the finite-$K$ computation. We distinguish: - $g_{\rm local}$: the local mixed derivative computed by automatic differentiation. - $\widehat g_K$: the reference hypergradient computed by differentiating through $K$ inner Adam updates, initialized with the snapshot’s optimizer state. For an estimate $\widetilde g$ and a specified reference $g_{\rm ref}$, **Appendix G.5, Eqs. (58)–(59)** defines \$ \\text{Cosine}(\\tilde{g}, g_{\\text{ref}}) = \\frac{\\tilde{g}^\\top g_{\\text{ref}}}{\\Vert \\tilde{g} \\Vert_2 \\Vert g_{\\text{ref}} \\Vert_2} \$ Values closer to **1** indicate stronger directional agreement between the finite-difference estimate and the local reference gradient. Local directional agreement: **Fig. 11(a)**. The reference in this panel is $g_{\rm local}$. | $\epsilon$ | Epoch 0 | Epoch 20 | Epoch 40 | Epoch 45 | Epoch 50 | |---|---:|---:|---:|---:|---:| | **0.001 (training)** | 0.9615 | 0.9811 | 0.9136 | 0.9999 | 0.9975 | | 0.01 | 0.9769 | 1.0000 | 0.9977 | 1.0000 | 1.0000 | | 0.1 | 0.9998 | 1.0000 | 0.9988 | 1.0000 | 0.9998 | These measurements support that **the centered finite-difference estimator shows strong directional agreement with its local mixed-derivative reference**
Reviewer_mtnyOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2(+1)
Read review
Paper summary

META targets two failure modes of distribution alignment that appear when a classifier trained on one time-series domain is applied to another. Joint MMD constrains only class-conditional means, so a mean-preserving drift can inflate intra-class spread, while a global contrastive pull stretches each sub-mode along its radial direction. The method pairs time-frequency contrastive learning with MetaMMD, which learns a kernel bandwidth per pair through a bilevel objective, so cross-domain pairs inside a sub-mode attract while the rest fade. The paper argues that the contrastive optimum collapses classes and that sub-modal variance decays exponentially, and experiments cover eight datasets with per-pair accuracy and Macro-F1 tables in the appendix. The diagnosis is well posed and the design is economical, which makes the distance between the claims and the evidence the more visible.

Critical strengths
  1. Sharp Diagnosis: Intra-class and sub-modal dispersion are defined as measurable quantities, so the motivation can be tracked rather than asserted. Figure 4 tracks both across datasets, and Lemma 2.1 gives the class-mean blind spot a concrete form.
  2. Evaluation Breadth: The comparison spans eight datasets and a long baseline list that includes recent time-series methods. Accuracy and Macro-F1 are reported per transfer pair, so the aggregate claim can be checked against individual failures.
  3. Consistent Gains: META leads every dataset column of the main table, with its largest margin on Sleep-EDF at 8.2 points. The advantage appears across task families rather than in a single dataset or sensor setup.
Critical weaknesses
  1. The Theorems Rest on Assumptions That Do the Work: Theorem 4.1 appears unqualified in the main text, while the appendix proves it only after adding an assumption that positive pairs are already within ϵ\epsilon of each other. That assumption is an approximate version of the collapse the theorem asserts, so the exact conclusion does not follow for any ϵ>0\epsilon > 0. Proposition 3.1 has the opposite problem, since Assumption C.1 fixes the entire displacement field and is strictly stronger than the covariance inequality derived from it. Assumption D.1 fixes the sign of each pair's gradient, which under the stated dynamics is the sign of the weight derivative, so the lemma's content is that assumption restated.
  2. The Outer Objective Cannot Select Sub-Modal Bandwidths: The meta-model is trained to maximize a domain-discrimination loss indexed by domain label alone, so nothing in the outer objective rewards keeping a class's sub-modes apart. Bandwidths that merge sub-modes of the same class are therefore not penalised, which is the opposite of what the method claims to learn. Lemma D.1 then removes the remaining constraint, because once within-domain and cross-sub-mode weights decay to zero only a kernel attraction between matched cross-domain pairs survives. That attraction has no negatives, so it cannot preserve the separation the bandwidths are supposed to protect.
  3. The Ablation Cannot Separate the Components It Removes: The contrastive-loss column of the ablation table reproduces the w/o M. column on every dataset, and its delta row is computed from that duplicated mean rather than from the full model. Those deltas therefore describe a model that already lacks MetaMMD, not the contribution of each contrastive term. No column removes the frequency branch alone, so the benefit of a frequency view rests on a single dataset reported as point accuracies. The bandwidth study covers four of the eight datasets, which leaves half the evaluation untested for the module the paper treats as central.
  4. The Protocol Is Described Inconsistently: Table 1's caption advertises a standard deviation over domain pairs and five independent runs, but the values are the spread across pairs alone. Section 5.1 and Appendix F.3 both describe an across-pair spread, so the caption invites a reading of run stability that its own numbers do not support. The encoder description is split in the same way, since Appendix F.3 shares one 1D-CNN across the UDA baselines while time-series baselines keep their own extractors. It still concludes that differences are attributable solely to the adaptation algorithm, which makes the margins against EDEN and ACON a mix of algorithm and backbone.
Questions for authors
  1. Can Theorem 4.1 be recovered from the global optimum of the objective META actually minimizes, rather than from an assumption that positive pairs are already coincident?
  2. What in the outer objective discourages the bandwidth solution from merging same-class sub-modes, given that the loss is indexed by domain label only?
  3. What does MetaMMD contribute after within-domain and cross-sub-mode weights have vanished, and can a constant learned bandwidth be reported under the same bilevel training as a control?
  4. Suggestion: report run-level standard deviations at the dataset level, and correct Table 1's caption so the reported spread is identified as across domain pairs.
12 comments
AuthorsOpenReview ↗

Response to Reviewer mtny W1

We thank the reviewer for carefully examining the theoretical arguments and insightful comments. We have uploaded the revised manuscript, with all changes highlighted in blue. Below we address the concerns separately.

W1. The Theorems Rest on Assumptions That Do the Work. Theorem 4.1 appears unqualified in the main text, while the appendix proves it only after adding an assumption that positive pairs are already within ϵ\epsilon of each other. That assumption is an approximate version of the collapse the theorem asserts, so the exact conclusion does not follow for any ϵ\epsilon. Proposition 3.1 has the opposite problem, since Assumption C.1 fixes the entire displacement field and is strictly stronger than the covariance inequality derived from it. Assumption D.1 fixes the sign of each pair's gradient, which under the stated dynamics is the sign of the weight derivative, so the lemma's content is that assumption restated.

1. Theorem 4.1 — exact collapse remains the conclusion after removing the proximity assumption.

Thank you for highlighting the need to make the assumptions and scope explicit. We agree that the former condition $ \Vert \widetilde{z}_i - \widetilde{z}_p \Vert_2 \leq \epsilon$ was strong. We have removed this condition and replaced the alignment-based argument with a self-contained global-optimality proof in Appendix B, consistent with the supervised contrastive geometry characterized by [2] Graf et al., Dissecting Supervised Contrastive Learning, ICML 2021.

The revised Assumption B.1 specifies a fixed batch containing C≥2C\geq2 classes with m≥2m\geq2 views per class, unit-normalized features, and dimension D≥C−1D\geq C-1. We optimize the isolated SupCon loss over these feature configurations at a fixed temperature τ>0\tau>0. These conditions place no proximity constraint on positive pairs. The revised result still establishes that same-class normalized features collapse to a single class center, yielding Σc=0\boldsymbol\Sigma_c=\mathbf0 and Vin=0V_{\mathrm{in}}=0, while the class centers form an origin-centered regular simplex satisfying ∑cμc=0\sum_c\boldsymbol\mu_c=\mathbf0 and μc⊤μc′=−1/(C−1)\boldsymbol\mu_c^\top\boldsymbol\mu_{c'}=-1/(C-1) for c≠c′c\neq c'.

[2] F. Graf, C. Hofer, M. Niethammer, and R. Kwitt. "Dissecting Supervised Contrastive Learning." International Conference on Machine Learning (ICML), pp. 3821–3830, PMLR, 2021.

2. Proposition 3.1 — Assumption C.1 is realized by an update derived directly from the contrastive loss.

We thank the reviewer for prompting us to make the connection to the optimization objective explicit. Assumption C.1 is realizable in a valid target-domain configuration. The new Appendix C.2, Proposition C.2 constructs such a configuration and derives the stretching displacement from the loss gradient.

Specifically, consider two target samples from the same semantic sub-mode, each with aligned temporal and frequency views,

z_2^t=z_2^f=(\cos\theta,-\sin\theta).$$ Each sample’s other view is its positive, while the other sample’s views are negatives. The target instance-contrastive objective in Eq. (4) therefore reduces to $$L(\theta)= \log\left[1+2\exp\left(\frac{\cos(2\theta)-1}{\tau}\right)\right].$$ Direct differentiation gives $$L'(0)=0,\qquad L''(0)=-\frac{8}{3\tau}<0.$$ For the angular mapping $T_\eta$ induced by the unit-sphere gradient update, $$T_\eta(0)=0,\qquad T_\eta'(0)=1+\frac{2\eta}{3\tau}>1.$$ Thus, the collapsed configuration is an **unstable fixed point along the splitting direction**. Every sufficiently small positive perturbation is amplified, so $\theta^+=T_\eta(\theta)>\theta$. The centered displacement induced by this update satisfies $$r_i=\lambda(\theta)u_i,\qquad \lambda(\theta)=\frac{\sin\theta^+}{\sin\theta}-1>0.$$ **This is exactly Assumption C.1 with zero residual.** Moreover, $$\Sigma(\theta)= \begin{pmatrix} 0&0\\ 0&\sin^2\theta \end{pmatrix}, \qquad \|\Sigma(T_\eta(\theta))\|_F =\sin^2\theta^+>\sin^2\theta =\|\Sigma(\theta)\|_F.$$ The construction therefore establishes that an arbitrarily compact sub-mode can expand after an actual contrastive gradient update. **Proposition 3.1 and its accompanying discussion** state this possibility, with the covariance implication proved in **Appendix C.1** and its realization proved in **Appendix C.2**. This local result does not conflict with Theorem 4.1: Proposition 4.1 characterizes a global optimum of label-supervised source SupCon. Proposition C.2 analyzes a local target instance-contrastive update, in which different samples are negatives even when they belong to the same semantic sub-mode.
AuthorsOpenReview ↗

Response to Reviewer mtny W1

3. Assumption D.1 — the sign conditions follow from explicit local geometry, and the observed signs are measured from gradients.

The sign of qijq_{ij} is determined by the actual outer gradient and the MMD pair direction, qij=g~i⊤Δij,q_{ij}=\widetilde g_i^\top\Delta_{ij}, where Δij\Delta_{ij} denotes the pair displacement direction, with the positive learning-rate, normalization, and kernel-weight factors separated as in Eq. (40). We added Appendix D.2, to derive sufficient geometric conditions for the signs.

Let μ\mu be the matched source sub-mode center and assume the local confusion gradient admits

\kappa_i\geq\kappa>0,\qquad \|\xi_i\|\leq\bar\xi.$$ Writing $r_i=\|z_i-\mu\|$, the derivations yield the following bounds. | Pair group | Geometric condition | Bound and sufficient sign condition | | --------------------------------------------------------------------- | ----------------------------------------------------------------------- | --------------------------------------------------------------------- | | $S$, same-class cross-domain pairs within the same sub-mode | $\|z_{s,j}-\mu\|\leq\rho$, $r_i>\rho$ | $q_{ij}\geq\kappa r_i(r_i-\rho)-\bar\xi(r_i+\rho)>0$ | | $W$, within-domain target pairs satisfying the stated radial ordering | $r_j=\|z_j-\mu\|<r_i$ | $q_{ij}\leq-\kappa r_i(r_i-r_j)+\bar\xi(r_i+r_j)<0$ wh | | $X$, same-class cross-domain pairs across different sub-modes | $(\mu - z&#95;i)^\top (z'&#95;{s,j} - z&#95;i) \leq -\delta&#95;u < 0$ $\Vert z'&#95;{s,j} - z&#95;i\Vert \leq D$ | $q_{ij}\leq-\kappa\delta_u+\bar\xi D<0$ if $\bar\xi D<\kappa\delta_u$ | **Empirical evidence for observed signs** We compute the raw $q_{ij}$ from the learned model’s gradients and pair directions, then group the measurements using the fixed reference sub-mode identities. In **Fig. 12**, the observed agreement rates during late training on WISDM are | Group | Predicted sign | Observed agreement | |---|---|---:| | $S$ | $q_{ij}>0$ | 63.2% | | $X$ | $q_{ij}<0$ | 69.5% | | $W$ | $q_{ij}<0$ | 63.3% | **Table 9** additionally reports the following positive-sign rates for $S$, averaged over epochs and then equally over the available runs for each transfer. | Scenario | Fraction of $S$ pairs with $q_{ij}>0$ | |---|---:| | WISDM $26\to2$ | 84.60% | | HAR $12\to16$ | 69.85% | | HAR $2\to11$ | 91.04% | | WISDM $2\to32$ | 94.15% | These measurements show that the predicted sign patterns occur in the sampled training states. The rates below 100% also establish the empirical scope of the assumption.
AuthorsOpenReview ↗

Response to Reviewer mtny W2

W2. The Outer Objective Cannot Select Sub-Modal Bandwidths. The meta-model is trained to maximize a domain-discrimination loss indexed by domain label alone, so nothing in the outer objective rewards keeping a class's sub-modes apart. Bandwidths that merge sub-modes of the same class are therefore not penalised, which is the opposite of what the method claims to learn. Lemma D.1 then removes the remaining constraint, because once within-domain and cross-sub-mode weights decay to zero only a kernel attraction between matched cross-domain pairs survives. That attraction has no negatives, so it cannot preserve the separation the bandwidths are supposed to protect.

Response. We thank the reviewer for raising this question. The assertion that domain-only supervision cannot penalize sub-modal merging overlooks the local source–target coverage mismatch that a merge can create. Likewise, the conclusion that attraction without negatives cannot preserve separation overlooks the distinct source sub-modes to which the target groups are attracted. We have added empirical evidence to validate this effect. We clarify both points below.

1. Domain labels can detect the distributional consequence of merging.

Suppose two source sub-modes remain separated. If target sub-modes merge and leave one source region insufficiently covered, the local source–target imbalance makes domain classification easier. The discriminator can detect this imbalance using domain labels.

More formally, let H⋆H^\star be the minimum domain cross-entropy under balanced domain sampling, with the loss averaged over both domains and natural logarithms. For an unrestricted optimal discriminator, H⋆=log⁡2−JS⁡(Ps,Pt).H^\star=\log 2-\operatorname{JS}(P_s,P_t). If a source sub-mode region AA satisfies Ps(A)−Pt(A)≥δ>0P_s(A)-P_t(A)\geq\delta>0, then log⁡2−H⋆=JS⁡(Ps,Pt)≥δ22.\log 2-H^\star=\operatorname{JS}(P_s,P_t)\geq\frac{\delta^2}{2}. Thus, a merge that creates a coverage deficit incurs a quantifiable cost relative to complete domain confusion. This establishes how a domain-only objective can disfavor merging under separated-source and local-coverage conditions. The argument concerns the ideal discriminator objective, with practical behavior depending on the learned discriminator.

2. Lemma D.1 describes selective attraction toward distinct source sub-modes.

Under the geometric conditions detailed in Appendix D.2, Lemma D.1 connects the outer gradient to pair-specific interaction weights. For the Gaussian kernel, the interaction weight wij=exp⁡[−dij2/(2σij2)]σij2w_{ij}=\frac{\exp[-d_{ij}^{2}/(2\sigma_{ij}^{2})]}{\sigma_{ij}^{2}} is maximized at σij⋆=dij/2\sigma^\star_{ij}=d_{ij}/\sqrt2. The stated dynamics favor larger interaction weights for matched cross-domain pairs and smaller weights for mismatched pairs. Suppressing cross-sub-mode attraction removes a force that would otherwise mix different groups. The surviving matched attractions can preserve separation because they lead to different source anchors. To make this explicit, consider two target groups with centers μmt,μnt\mu_m^t,\mu_n^t and their matched source centers μms,μns\mu_m^s,\mu_n^s. Define Mm(t)M_m(t) as the mean squared distance of target group mm to μms\mu_m^s, as in Lemma D.2, Appendix D.1, Eq. (43)–(45). Jensen’s inequality and the triangle inequality give

\geq \|\mu_m^s-\mu_n^s\|-\sqrt{M_m(t)}-\sqrt{M_n(t)}.$$ Consequently, target centers remain separated whenever the source-center distance exceeds the two alignment errors. **Explicit repulsion between target groups is therefore unnecessary in this regime.** **3. The learned bandwidths exhibit the predicted differentiation.** We added the bandwidth trajectory on HAR $2\to11$ in **Section 5.4, Fig. 5(a)**. Groups $S$ and $X$ contain same-class cross-domain pairs from the same and different reference sub-modes, respectively, and $W$ contains within-domain pairs. We track $\sigma_{ij}/\sigma^\star_{ij}$, whose value of 1 maximizes the interaction weight at the current pair distance. Following joint training, the $S$ trajectory approaches the interaction-maximizing bandwidth more closely than $X$ and $W$. At epoch 50, its median distance from 1 is **0.1002**, compared with **0.1648** for $X$ and **0.7202** for $W$. The learned bandwidths thus differentiate the three pair groups in the direction predicted by the analysis. | Median normalized bandwidth | End of discriminator pretraining, epoch 20 | Joint training, epoch 50 | |---|---:|---:| | $S$ | 0.0882 | 1.1002 | | $X$ | 0.0978 | 0.8352 | | $W$ | 0.4380 | 0.2798 |
Reviewer_F6jTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes META, a time-series unsupervised domain adaptation framework that combines Time-Frequency Contrastive Learning to improve intra-class compactness and MetaMMD to preserve sub-modal structure through bi-level bandwidth optimization. The authors provide theoretical analyses and evaluate the method on eight datasets across four classification tasks. Experimental results, ablation studies, and feature dispersion analyses demonstrate its effectiveness.

Critical strengths

1.The paper identifies intra-class dispersion and sub-modal structure degradation as two important challenges in time-series domain adaptation and proposes a unified framework to address them.

2.META achieves competitive results across eight datasets and four classification tasks. The ablation studies and feature dispersion analyses further demonstrate the effectiveness of the proposed components.

Critical weaknesses
  1. The paper employs supervised contrastive learning to pull same-class samples closer together, with the theoretical analysis emphasizing that same-class features eventually collapse toward a single class center. Meanwhile, MetaMMD aims to preserve distinct sub-modes within each class and prevent them from merging. These two objectives may potentially conflict, as continuous intra-class feature contraction could weaken the original sub-modal structure. Although the paper separately validates the effectiveness of both modules, it does not sufficiently explain how they are coordinated during joint optimization. The authors are encouraged to further clarify the relationship between intra-class compactness and sub-modal preservation, and how MetaMMD mitigates the potential impact of intra-class contraction on sub-modal structure.
  2. The paper introduces MetaMMD to reduce sub-modal dispersion while preserving the structural independence of different sub-modes. Figure 4(b) shows a reduction in VsubV_{sub}, but this metric primarily measures within-sub-mode compactness and does not directly reflect the separation between different sub-modes. The authors are encouraged to clarify how sub-modes are identified in the experiments and to what extent the current results support the claim of multi-modal structure preservation.
  3. Section 5.3 refers to the baseline in Figure 4 as “HoMM (JMMD),” whereas the figure legend labels it as “JMMD.” Since HoMM and JMMD are two distinct distribution alignment methods, the authors should clarify which baseline was actually used and ensure consistent terminology to avoid ambiguity in interpreting the experimental results.
Questions for authors

1.How does META balance intra-class compactness and sub-modal preservation during joint optimization, given that supervised contrastive learning encourages same-class feature collapse while MetaMMD aims to preserve distinct sub-modes? 2.How are sub-modes identified when computing VsubV_{sub} in Figure 4(b), and how do the reported results support the claim that different sub-modal structures are preserved rather than merely becoming more compact?

6 comments
AuthorsOpenReview ↗

Response to Reviewer F6jT W1

We thank the reviewer for recognizing the importance of distinguishing intra-class dispersion from sub-modal structure, and for the constructive comments on our analysis. We have uploaded the revised manuscript, with all changes highlighted in blue. Below we provide our detailed responses.

W1. The paper employs supervised contrastive learning to pull same-class samples closer together, with the theoretical analysis emphasizing that same-class features eventually collapse toward a single class center. Meanwhile, MetaMMD aims to preserve distinct sub-modes within each class and prevent them from merging. These two objectives may potentially conflict, as continuous intra-class feature contraction could weaken the original sub-modal structure. Although the paper separately validates the effectiveness of both modules, it does not sufficiently explain how they are coordinated during joint optimization. The authors are encouraged to further clarify the relationship between intra-class compactness and sub-modal preservation, and how MetaMMD mitigates the potential impact of intra-class contraction on sub-modal structure.

Response. We thank the reviewer for identifying this tension. We clarify that the theoretical collapse describes the isolated optimum of source-domain SupCon, which does not imply target sub-mode collapse under joint optimization. We have clarified in Section 4.1 and Appendix B that the collapse result characterizes a global optimum of the isolated, class-balanced source SupCon loss on a fixed batch. In the joint objective, SupCon promotes source class compactness, while MetaMMD learns pair-specific bandwidths that can favor alignment with the corresponding source sub-mode over attraction toward a different one (Section 4.2 and Appendix D.2). This suggests a mechanism for selective alignment. Whether it actually prevents merging is a separate empirical question, which we address below.

1. Measuring compactness and separation separately (Sections 5.3–5.4; Appendix E.4).

We revised the dispersion evaluation in VinV_{\mathrm{in}} and VsubV_{\mathrm{sub}} . Rather than fitting a GMM to each learned representation, we use a fixed external GMM partition shared across all methods. VinV_{\mathrm{in}} measures class-level dispersion, while VsubV_{\mathrm{sub}} measures dispersion within these fixed groups.

To more comprehensively evaluate the sub-modal structure, we introduced additional metrics in Appendix E.4: Dmin⁡D_{\min} (the minimum distance between two same-class sub-mode centers), RsubR_{\rm sub} (that distance relative to within-sub-mode spread), Silhouette (within-group cohesion versus separation), Submode Match (agreement with the nearest same-class source sub-mode center), and Merge Rate (the fraction of same-class reference pairs assigned to the same dominant non-noise cluster).

2. Does lower dispersion come at the cost of separation? (Fig. 4 and Table 4).

The table gives the values plotted in Fig. 4. Each entry is Vin↓V_{\rm in}\downarrow / Vsub↓V_{\rm sub}\downarrow / Rsub↑R_{\rm sub}\uparrow.

DatasetSourceJANw/o MetaMMDMETA
HHAR_SA0.0386 / 0.0257 / 0.31040.0541 / 0.0563 / 0.27820.0370 / 0.0296 / 0.32220.0264 / 0.0261 / 0.3439
WISDM0.0222 / 0.0331 / 0.53200.0366 / 0.0506 / 0.26310.0301 / 0.0314 / 0.30620.0081 / 0.0184 / 0.3527
HAR0.0286 / 0.0398 / 0.16680.0561 / 0.0568 / 0.26090.0372 / 0.0288 / 0.37840.0270 / 0.0239 / 0.3604
Sleep-EDF (EEG)0.1105 / 0.0602 / 0.48250.0647 / 0.0265 / 0.43760.0533 / 0.0210 / 0.56430.0106 / 0.0049 / 0.5886

Table 4 demonstrates that even as the reference groups become tighter, their closest centers move farther apart, Sub-mode Match increases, and the Merge Rate drops.

EMG 1→31\to3Vin↓V_{\rm in}\downarrowVsub↓V_{\rm sub}\downarrowDmin⁡↑D_{\min}\uparrowRsub↑R_{\rm sub}\uparrowSilhouette ↑Submode Match ↑Merge Rate ↓
w/o MetaMMD0.12890.08400.00830.0256−0.028943.43%56.60%
META0.09070.07270.04780.15430.038658.49%18.87%
AuthorsOpenReview ↗

Response to Reviewer F6jT W1

3. What happens during training and as the loss balance changes? (Fig. 5b–c). Fig. 5b tracks DinterD_{\rm inter}, the minimum same-class sub-mode center distance. Selected epochs are shown below (means over three seeds; discriminator pretraining ends at epoch 20).

Dinter↑D_{\rm inter}\uparrowEpoch 0Epoch 20Epoch 24Epoch 50Epoch 70
w/o MetaMMD0.02900.01550.00340.01050.0174
META0.02900.01550.00400.01480.0276

Fig. 5c varies ρ=λSCon/λTCon\rho=\lambda_{\rm SCon}/\lambda_{\rm TCon} while keeping the sum of these two weights fixed. The table gives all five final configurations for the plotted seed. At ρ=0.5\rho=0.5, META achieves both lower dispersion and greater separation than the ablation at the same loss balance. At ρ=0\rho=0, the ablation attains the smallest VsubV_{\rm sub} in the sweep, but its DinterD_{\rm inter} is only 0.0082.

ρ\rhow/o MetaMMD: Vsub↓V_{\rm sub}\downarroww/o MetaMMD: Dinter↑D_{\rm inter}\uparrowMETA: Vsub↓V_{\rm sub}\downarrowMETA: Dinter↑D_{\rm inter}\uparrow
00.03130.00820.04070.0092
0.250.04440.00780.03450.0221
0.50.03860.01520.03760.0258
10.04390.00930.04060.0105
∞\infty (w/o TCon)0.03840.00700.03480.0103
AuthorsOpenReview ↗

Response to Reviewer F6jT W2

W2. The paper introduces MetaMMD to reduce sub-modal dispersion while preserving the structural independence of different sub-modes. Figure 4(b) shows a reduction in VsubV_{\mathrm{sub}}, but this metric primarily measures within-sub-mode compactness and does not directly reflect the separation between different sub-modes. The authors are encouraged to clarify how sub-modes are identified in the experiments and to what extent the current results support the claim of multi-modal structure preservation.

Response. Thank you for pointing out this distinction. Two sub-modes can each have low VsubV_{\mathrm{sub}} even if their centers coincide. We therefore added two checks that address your concern directly. We measure the distance between fixed reference groups, and we trace whether samples from different groups enter the same cluster when each method’s embeddings are independently re-clustered.

How we identify and track sub-modes (Section 5.4, Fig. 6). In the Sankey diagrams (Fig. 6), the left nodes are the fixed reference groups, the right nodes are the HDBSCAN clusters, and each ribbon counts samples. Adding MetaMMD reduces merged reference pairs from 3/10 to 0/10 for HAR 12→1612\to16 (class 5) and from 1/3 to 0/3 for WISDM 18→2018\to20 (class 5).

On HHAR_SA 1→6, the same 483 samples are tracked across three methods in Appendix G.2, Fig. 8. Without MetaMMD, 88/88 R3 samples and 68/110 R4 samples enter C1. With META, 87/88 R3 samples enter C3, while 73/110 R4 samples enter C1. Across the five reference groups, merged pairs fall from 10/10 to 3/10. Compared with the shared-bandwidth variant BSB-MMD, META also reduces the number of noise assignments from 335/483 (69.4%) to 137/483 (28.4%). For R4 alone, the count falls from 98/110 to 37/110.

How far the evidence extends (Table 4 and Appendix G.3, Fig. 9). On EMG 1→3, META reduces VsubV_{\mathrm{sub}} from 0.0840 to 0.0727 while the minimum same-class center distance increases from 0.0083 to 0.0478 and the merge rate falls from 56.60% to 18.87%. Here, both center distances and sample assignments support the interpretation that the groups become more compact while remaining distinct. Figure 9 reports the following pooled, all-class merge rates (%) across transfer scenarios. We therefore interpret the results as evidence that MetaMMD can preserve separation between identifiable reference groups and reduce merging on average over the reported scenarios.

WISDM26→221→317→3020→302→3212→725→2912→1918→204→15Mean
w/o MetaMMD18.1815.0013.8912.5010.2610.007.504.764.760.009.69
META6.8215.002.784.1710.260.000.004.7614.290.005.81
EMG1→31→20→30→20→13→12→32→12→0Mean
w/o MetaMMD56.6052.9442.1125.0021.6216.228.168.117.3226.45
META18.8723.532.6319.6410.8116.228.1616.2221.9515.34

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.