acceptodds
Under review as a conference paper at ICLR 2027

SCALE: Support Calibration with Anchored Local Experts for Imbalanced Time-Series Classification

Abstract

Rich frozen time-series representations provide many feature directions, but scarce minority labels may be insufficient to identify those that generalize. Our diagnostics reveal a gap between training and test recall for the minority class under class-balanced readouts. With the encoder held fixed, readouts trained with more labeled examples achieve higher ROC-AUC. Thus, readout performance under sparse supervision can fall short of the discrimination supported by the same frozen representations. Motivated by these observations, we propose Support Calibration with Anchored Local Experts (SCALE) to improve minority generalization without additional labels or encoder updates. SCALE uses shared local references to construct compact residual coordinates and fits complementary experts on observed and modeled support. These experts propose minority-directed corrections to a class-balanced anchor, with training cross-validation selecting the correction rule. Across 690 univariate encoder–task pairs, SCALE achieves the highest aggregate balanced accuracy, Macro-F1, and Min-F1 among the compared methods. It improves on cRT by balanced-accuracy and Min-F1 percentage points, with larger gains under more severe imbalance in a matched cohort. Multivariate evaluation also shows positive aggregate gains over cRT.

Primary area:
learning on time series and dynamical systems
Keywords:
imbalanced time series classification, representation learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_cARAOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper addresses imbalanced time-series classification with frozen representations. In diagnostics on 13 binary datasets and four pretrained encoders, the authors show that minority training recall rises while minority test recall falls as minority support shrinks. With the encoder held fixed, a readout trained with more labels achieves higher ROC-AUC. They conclude that the frozen features contain more class information than a readout can exploit under scarce minority supervision.

They propose SCALE:

  1. LRG: Features are standardized and L2-normalized, then clustered with k-means. Each sample gets a soft local reference and scale from the cluster centers; the signed residual to this reference is projected with shrinkage LDA, and the resulting compact coordinate u is appended to the features.
  2. SSE: Synthetic minority samples are drawn from a regularized low-rank Gaussian.
  3. Correction: A class-weighted logistic anchor provides the default prediction. Three experts (logistic on z, u, an RBF-SVM on observed data, and an RBF-SVM on observed plus synthetic data) can switch a prediction to the minority class if at least q of them vote for it, with q selected by training CV.

SCALE is evaluated on 690 univariate and 160 multivariate encoder–task pairs and achieves the best aggregate balanced accuracy, Macro-F1 and Min-F1.

Critical strengths
  1. Large and careful evaluation: All methods use the same frozen features, splits and tuning budget. Statistics are computed at the dataset level with Holm correction, and all 208 diagnostic conditions are reported, including negative ones (Table 20).
  2. Transparent recovery–harm analysis: Proposition 1 and Figure 4 show that the growing gain under imbalance comes from increased minority recovery while harm stays roughly constant. This explains how the correction improves balanced accuracy, rather than only reporting that it does.
Critical weaknesses
  1. It is unclear where the gains come from and under which conditions they hold.
    1. Imbalance ratio and absolute minority count are confounded, because only the minority class is downsampled. In the Figure 1 example there are only 93 minority samples, and many univariate tasks have just 15–40 at 20:1. In the multivariate benchmark, where several datasets have much larger minorities (e.g. 925 or 534 at 20:1), the gains are smaller (0.69 BA vs. 2.06 univariate), and the Macro-F1 and Min-F1 gains are not significant (Holm-adjusted p = 0.11 and 0.15). Although the datasets differ, this suggests that the claimed dependence on imbalance may partly be a dependence on the absolute number of minority samples.
    2. The experts are never evaluated on their own: Table 12 always includes the anchor, and even the raw-ensemble baseline is anchor plus RBF expert. It is therefore unclear whether the one-way gating is needed or whether a strong RBF-SVM alone explains the gains.
    3. Tables 1–3 report only means, without standard deviations or confidence intervals.
  2. The added complexity is substantial relative to the gain. SCALE requires about 3.2× the mean runtime of logit adjustment, the strongest standard baseline (187.6 s vs. 58.0 s per run, Table 16; about 11× by median). It also requires an eight-parameter search, four classifiers and RBF kernels whose cost grows quadratically with the (synthetically enlarged) training set. In return, it improves on logit adjustment by 1.34 BA points univariate and 0.66 BA points multivariate. It is unclear whether this complexity is justified.
Questions for authors
  • Clarification: Can you report standard deviations or confidence intervals across seeds and datasets for Tables 1–3?
  • Clarification: How do the experts L, O and E perform as standalone classifiers, and how does a plain majority vote over all four classifiers compare to the one-way correction?
  • Future work: SCALE operates on frozen embedding vectors, nothing in LRG, SSE or the experts uses temporal structure. Apart from the chosen datasets and encoders, nothing in this work is specific to time series, have you tested SCALE also on other domains?
  • Clarification: How does SCALE behave at a fixed ratio (e.g. 20:1) with a large absolute minority count (e.g. at least 1,000 minority samples)? Can you report per-dataset multivariate results, so that datasets with large minorities (FordChallenge, STEW) can be compared with those with very small ones (LowCost, UCDHE-Rowing)?
2 comments

W1.1 / Q4 — How does SCALE behave at a fixed ratio with a large minority count?

At 20:1, none of the eight multivariate datasets retained 1,000 minority training examples. On the two largest supports, SCALE improved balanced accuracy over cRT by 2.00 points on FordChallenge (925 minority examples) and 2.35 points on STEW (534). At the same ratio, the gains were −0.07 points on LowCost (15) and +1.30 on UCDHE-Rowing-MC (20). At 10:1, FordChallenge and STEW both exceeded 1,000 minority examples, with gains of 1.12 and 1.54 points, respectively.

We agree that our downsampling protocol varies the imbalance ratio and absolute minority count together. These per-dataset comparisons show that gains can persist with larger minority supports. The separate effects of ratio and count remain open. See Appendix H.4, Tables 26–27.

W1.2 / Q2 — How do the standalone experts and plain voting compare with one-way correction?

Thank you for asking how the one-way correction compares with the individual experts and a plain vote. The standalone RBF-SVM comparison is in our response to Reviewer 78zM (W1 / Q1). On the univariate benchmark, L, O, and E individually achieved 75.90%, 73.64%, and 75.90% balanced accuracy, respectively.

We independently tuned a plain vote over the same four classifiers. Each classifier cast one vote, the class with the most votes won, and the anchor broke ties when its prediction was among the tied classes. SCALE improved balanced accuracy from 76.20% with plain voting to 77.38%, a paired gain of 1.18 percentage points (95% CI [0.83, 1.54]). At 20:1, SCALE recovered more minority cases, with lower minority precision and majority recall. See Appendix E.4, Tables 16–17.

W1.3 / Q1 — Can you report uncertainty for Tables 1–3?

Yes. We added 95% source-dataset bootstrap confidence intervals for Tables 1–3. These intervals capture variation across source datasets with the support splits held fixed. We report variation across fitting seeds separately. For SCALE on the univariate benchmark, the SDs across the ten seed averages were 0.09, 0.06, and 0.14 percentage points for balanced accuracy, Macro-F1, and Min-F1, respectively. See Appendix G.2, Tables 21–23.

W2 — Is the additional fitting cost justified?

The cost is substantial. With cached embeddings, SCALE’s full search, refit, and test prediction took about 3.2× as long on average and 11× as long at the median as logit adjustment. The corresponding mean balanced-accuracy gains were 1.34 points on the univariate benchmark and 0.66 points on the multivariate benchmark. The trade-off matters most when a readout needs frequent refitting. When the fitted readout can be reused, the search cost is incurred once per task. See Table 20 for timings and Tables 1–2 for the accuracy results.

Q3 — Have you tested SCALE outside time series?

Our experiments to date cover time-series classification only. SCALE operates on frozen embeddings without direct access to the time axis, so it can in principle be applied to embeddings from other domains. We identify testing that applicability as future work in the revised conclusion.

I thank the authors for their replies. Based on this and the discussion with the others reviewers, I will keep my score.

Reviewer_c7APOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies imbalanced classification using frozen time-series representations. SCALE combines a class-balanced logistic anchor with three experts: a logistic classifier augmented by a supervised local-residual coordinate, an RBF-SVM classifier on observed support, and an RBF-SVM classifier trained with additional low-rank Gaussian minority samples. Cross-validation selects the configuration and the number of expert votes required to change an anchor prediction to the designated minority class, including an anchor-only option. The paper reports improvements across five representations, 58 univariate source datasets and eight multivariate datasets, together with component ablations and a decomposition of the corrections into minority recovery and harm to other classes.

Critical strengths

A practically relevant problem with consistent empirical gains: Figuring out how to improve minority classification without having to update a massive encoder is a very useful, real-world challenge. SCALE delivers solid results here: on the univariate benchmark, it improves balanced accuracy by 2.06 percentage points over cRT, 1.34 over logit adjustment, and 1.41 over the raw ensemble. It consistently comes out ahead across all five representations (Table 1). The multivariate version also shows overall improvement, though the evidence there is a bit softer. Rigorous and well-designed experiments: The evaluation setup is well put together. The methods properly share feature caches and data splits, and downstream steps (like preprocessing and classifiers) are correctly refitted inside the training folds. I also appreciated that the main ablations were tuned independently to ensure fairness. On the statistical side, the tests correctly aggregate within source datasets instead of artificially treating different encoders, ratios, or seeds as independent data points. Finally, using a matched cohort to study the imbalance trend was a smart way to keep the underlying data composition stable.

Critical weaknesses

W1. The benefit of local references is not separated from that of supervised feature projection. LRG combines normalization, local residual construction and a label-fitted LDA projection. Table 8 compares constructions that differ in supervision and dimension, while removing LRG in Section 4.4 also removes the linear expert. These comparisons support the complete construction, but do not isolate the contribution of local references. Appending a global shrinkage-LDA coordinate can already change a regularized classifier. A matched global-projection control is therefore needed. The K=1K=1 equivalence checked under the binary conditions in the audit appendix motivates this comparison; it does not show that the proposed local geometry is ineffective.

W2. The added value of the full correction system over simpler nonlinear alternatives remains unclear. Most standard baselines use logistic regression, whereas SCALE includes two RBF-SVM experts. The raw ensemble is an important nonlinear control, and SCALE's 1.41-point BA improvement over it is meaningful evidence. Nevertheless, that baseline still uses one-way correction of an anchor. It does not establish performance relative to a standalone nonlinear classifier that makes unrestricted predictions, or to a similarly tuned nonlinear classifier with conventional minority augmentation. For SSE, the \texttt{original} fitting mode also changes more than the positions of the minority samples: expansion changes their total loss contribution and the data-dependent RBF bandwidth (Appendix D). The balanced-reference mode controls these effects, but its results are not isolated. Therefore, the current evidence does not fully separate gains from the modeled minority geometry from gains due to reweighting or bandwidth changes. The independently retuned ablations support incremental benefits of both components; the additional gain of full SCALE over either single-component removal is approximately 0.4--0.5 BA points (Table 3).

W3. Robustness to the choice of scarce minority observations is untested. The ten seeds vary downstream CV, fitting and support generation, but retain the same downsampled minority observations (Appendix H.1, lines 1227--1233). This is disclosed and makes the fixed-benchmark comparisons valid. However, it does not establish whether the gains persist when a different small set of minority examples is available. That uncertainty is especially relevant here because both the discriminative projection and the Gaussian support model are estimated from those examples, with retained minority counts as low as 15.

Questions for authors

Q1. Does SCALE retain a consistent advantage when its local-residual coordinate is compared with a global shrinkage-LDA coordinate computed directly from z~\tilde z? Please keep the shrinkage estimator, projection dimension and scaling, expert structure, and training-only selection protocol identical. A K=1K=1 variant would provide a complementary check. This comparison would establish whether local references add value beyond supervised projection alone.

Q2a.How does SCALE compare with a directly tuned class-weighted RBF-SVM, and with an RBF-SVM using conventional minority augmentation such as SMOTE, on the same feature cache? Please use matched preprocessing and comparable regularization and bandwidth searches. In particular, clarify whether row normalization is identical for SCALE and the raw ensemble. These comparisons would establish the practical value of the additional components and the restricted correction rule.

Q2b. Does SSE still improve performance when the total loss weight per class and the kernel bandwidth are held fixed between the observed-support and expanded-support classifiers? Reporting the corresponding comparison in the balanced-reference mode would help isolate the contribution of the generated support.

Q3. On a representative subset of the pretrained-encoder tasks, do the gains over cRT and the raw ensemble persist across independently sampled minority training subsets? Please share each sampled subset across methods, repeat downstream selection using training data only, and report the paired gains, their variability and the frequency of deterioration. A modest study using cached embeddings would directly address this concern without retraining encoders or repeating the full benchmark.

3 comments

W1 / Q1 — Separating locality from supervised projection

Thank you for pointing out that a global supervised coordinate could itself improve the readout. We describe the matched experiment in our response to Reviewer 78zM (W1 / Q2). With K = 1, every example uses the same centre and scale, so shrinkage LDA yields a global supervised coordinate on the row-normalised embeddings. Both variants use the same normalisation, so the comparison tests what example-dependent references and scales add.

The overall balanced-accuracy gain was positive in every encoder and seed average, with a paired 95% confidence interval of [0.07, 0.20] percentage points. At 20:1, local references also improved Min-F1 by 0.53 points. See Appendix B.5, Table 10.

W2 / Q2a — Standalone RBF-SVM and SMOTE-RBF

We describe the shared standalone RBF-SVM experiment in our response to Reviewer 78zM (W1 / Q1). To address your question about conventional minority augmentation, we also evaluated SMOTE-RBF. The two standalone baselines searched the same regularisation and bandwidth ranges, with SMOTE-RBF also tuning the neighbour count.

Both baselines and the RBF branches of SCALE and the raw ensemble use the same base preprocessing. We fit column standardisation within each fitting partition and then apply row-wise L2 normalisation. SCALE improved univariate balanced accuracy by 1.63 percentage points over SMOTE-RBF (paired 95% CI [1.03, 2.23]). See Appendix C.1 and Tables 1–2.

W2 / Q2b — Testing synthetic support with loss weights and bandwidth fixed

Thank you for pointing out that the original fitting mode changes class loss weights and bandwidth when synthetic samples are added. We used balanced-reference fitting to hold the feature coordinates, regularisation, total loss weight per class, and bandwidth fixed between the observed and expanded experts.

Adding synthetic support improved univariate balanced accuracy by 1.23 percentage points for the expanded expert. We also held the expanded expert’s place in the gate and the vote threshold fixed, changing only whether that expert used synthetic support. This improved the final SCALE prediction by 0.32 points. See Appendix E.3, Table 15.

W3 / Q3 — Sensitivity to the choice of minority examples

We describe the shared ten-draw resampling protocol in our response to Reviewer 78zM (W2 / Q3). Across all four ratios, SCALE’s balanced-accuracy gains over both cRT and the raw ensemble remained statistically significant in paired tests on source-dataset mean differences (Holm-adjusted p < 0.006).

Against the raw ensemble, the resampled gains averaged 2.27 ± 0.26 points at 10:1 and 3.71 ± 0.34 points at 20:1, close to the fixed-support gains of 2.60 and 3.85 points. At 20:1, SCALE scored below cRT in 32 of 520 paired conditions (6.2%) and below the raw ensemble in 45 of 520 (8.7%). Higher minority recall outweighed lower majority recall on average. See Appendix H.5, Tables 28–29.

We have updated the anonymous repository with the new experiments.

I thank the authors for their detailed answers and further experiments. They adress my major concerns and make the empirical contribution of the paper stronger. I thus raise my recommendation weak accept.

Q1: The new experiment with K=1 provides the global control and confirms the positive contribution of local references. The average contribution is small but increases under sever imbalance. Q2a: The new experiments with RBF-SVM and SMOTE-RBF baselines, which use matching preprocessing and comparable hyperparameters tuning procedures, show the benefits of the proposed full approach compared to simpler non-linear approaches on the univariate benchmark. Q2b: The new experiments isolate the impact of the generated support from changes in the class loss weights and the bandwidth. The fixed-gate experiment additionally shows that this positive impact persists in the final prediction. Q3: The support-resampling study confirms that the benefits hold when considering different minority training samples from the considered cohort. Sharing sampling draws across methods, repeating the training-only procedure, and measuring the deterioration frequency directly address the robustness issue.

My remaining qualifications concerns the gain from locality that still appears minor, and the multivariate analysis against SMOTE-RBF is less conclusive. Nevertheless, the developed approach brings consistent improvements under an extensive evaluation. The extra experiments address my principal worries.

Thank you for carefully reviewing the additional experiments. We appreciate your recognition of the consistent improvements across the evaluation. The K=1K=1 control retains the supervised projection, experts, support expansion, and gate. Local references provide an additional benefit within this framework, particularly under severe imbalance. We will state the overall comparison explicitly in Section 4.4. We also acknowledge the greater uncertainty in the multivariate comparison with SMOTE-RBF. Thank you again for reviewing our paper.

Reviewer_78zMOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies imbalanced time-series classification using five frozen representations: four pretrained encoders and a task-trained H-Inception encoder. Motivated by diagnostics of minority generalization under sparse training support, SCALE combines a class-balanced logistic anchor with three experts. The linear expert uses the original features augmented with Local Reference Geometry (LRG) coordinates. Both RBF-SVM experts use normalized features augmented with weighted LRG coordinates: one is trained on observed support, and the other on observed support augmented with generated minority samples. LRG applies a supervised projection to locally scaled residuals relative to soft combinations of k-means reference centers. Spectral Support Expansion (SSE) fits a low-rank Gaussian model with OAS covariance regularization in the same augmented coordinate used by the RBF experts. Expert votes can override the anchor only toward the designated minority class, with the vote threshold selected through training cross-validation and an anchor-only fallback available.

Across 690 univariate encoder–task pairs, SCALE improves over cRT by 2.06 balanced-accuracy points and 3.43 Min-F1 points, and over the raw-feature anchored ensemble by 1.41 balanced-accuracy points (Table 1). Multivariate experiments show smaller positive aggregate gains (Table 2). Proposition 1 and the accompanying diagnostics decompose changes in balanced accuracy into minority recovery and harm to previously correct predictions.

Critical strengths
  1. Extensive evaluation under a careful selection and statistical protocol. The univariate evaluation covers 138 tasks from 58 source datasets, five representations, and ten fitting/search seeds. Methods receive the same 100-trial TPE budget, with downstream transformations fitted within training folds and selection based on pooled out-of-fold predictions. Test performance is excluded from selection and tie-breaking. The statistical comparisons aggregate paired differences within source datasets before applying Wilcoxon tests and Holm correction, appropriately accounting for the dependence between encoder and imbalance-ratio views of the same dataset (§4.1, Appendix D). The baseline coverage includes logit adjustment, P2P, oversampling methods, and a raw-feature anchored ensemble that provides a useful nonlinear correction control. SCALE achieves the highest aggregate balanced accuracy for every evaluated representation. Its gains over cRT, logit adjustment, and the raw ensemble are significant on all three reported metrics: Holm-adjusted p < 10⁻⁵ against cRT and p < 0.001 against each of the latter two methods (§4.1). These results provide substantial evidence that the complete pipeline is useful across different frozen representations.

  2. An interpretable correction rule supported by informative component and error analyses. The one-way correction rule makes the permitted changes to the anchor explicit, and cross-validation can select the anchor-only fallback. Proposition 1 provides an exact decomposition of the resulting balanced-accuracy change. The empirical analysis uses this decomposition to show that, from 3:1 to 20:1 imbalance, the recovery contribution increases from 3.26 to 6.27 BA points while average harm remains between 1.94 and 2.23 points (Figure 4). Larger gains under scarcity are therefore associated with increased minority recovery while harm remains comparatively stable. The independently retuned ablations in Table 3 also support the practical contribution of both LRG and SSE within the selected pipeline. Together, these analyses make the results more informative than an aggregate accuracy comparison alone.

Critical weaknesses
  1. The contribution of the proposed architecture and local coordinates is not fully separated from simpler nonlinear alternatives. The raw-feature ensemble is a useful control, but it still applies the anchored one-way correction rule. It does not establish how SCALE compares with a standalone RBF-SVM independently selected for its own predictive performance. Since nonlinear experts are central to SCALE, this comparison would clarify the benefit of the full correction architecture relative to a simpler nonlinear readout. The existing LRG controls leave a more specific attribution question. With K = 1, the reference and scale become constant, so the added coordinate is a global supervised affine projection of the row-normalized representation. This is not equivalent to deleting LRG, since that representation is generally nonlinear in the original features z. Comparing this variant with SCALE’s input-dependent references and scales, while retaining the expert structure, projection procedure, and support expansion, would clarify what the local transformation contributes beyond a global supervised coordinate. Table 3 establishes component utility within the pipeline but does not address this narrower question.

  2. Robustness to support resampling and applicability to several rare classes remain untested. Each dataset–ratio task uses one fixed minority subset, shared across methods and encoders. The ten fitting/search seeds quantify variation conditional on that subset; they do not measure sensitivity to which minority observations are retained (Appendix H.1). This distinction matters for the scarcity analysis. For example, at 20:1, FloodModeling1_disc and AconityMINIPrinterSmall retain only 17 and 21 minority observations, respectively (Table 17). The matched cohort controls for changes in dataset composition across imbalance ratios, but the robustness of the gains to alternative support draws remains untested. The task construction also reduces one designated class while retaining all observations from the remaining classes, and the correction rule permits overrides only toward that designated class. The results establish effectiveness for this specific form of imbalance, with less evidence for settings requiring simultaneous corrections toward several rare classes.

Questions for authors

Q1. How does SCALE compare with a standalone RBF-SVM on the frozen features, with class-balanced loss weighting included in its search space? Please use the same 100-trial TPE budget and training folds, selecting regularization, bandwidth, and weighting using the standalone classifier’s own pooled OOF balanced accuracy. A comparison on a representative existing cohort would be sufficient.

Q2. On the same cohort, could the authors compare SCALE with a K = 1 variant that retains the supervised projection, support expansion, and expert structure, with the remaining parameters selected under the same training-only protocol? SSE should be refitted in the resulting coordinate, as in the main method. This would test the value of input-dependent local references and scales relative to a global supervised coordinate and its induced RBF metric.

Q3. For a small, pre-specified subset of the matched cohort at 10:1 and 20:1, how stable are SCALE’s gains over cRT and logit adjustment across several independent minority-downsampling draws? Using the same support draw across methods and reporting the distribution of paired gains would help establish whether the scarcity results are robust to the identities of the retained minority observations.

3 comments

W1 / Q1 — Standalone RBF-SVM

Thank you for raising this point. Our raw-feature ensemble retained the one-way gate, so it left open how SCALE would compare with a standalone RBF-SVM. Reviewer c7AP (W2 / Q2a) asked for the same comparison.

We evaluated a standalone RBF-SVM across the full univariate and multivariate benchmarks using the same frozen embeddings, training folds, and 100-trial TPE budget. We fixed class-balanced loss weights within each fitting fold and tuned regularisation and bandwidth using the SVM’s own pooled out-of-fold balanced accuracy. SCALE achieved higher mean balanced accuracy by 4.14 percentage points on the univariate benchmark and 2.46 points on the multivariate benchmark. See Appendix C.1 and Tables 1–2.

W1 / Q2 — Comparing local and global references

Thank you for suggesting this control. Reviewer c7AP (W1 / Q1) raised the same concern. Our original ablations showed that LRG improves the full pipeline, but we did not evaluate how much of that gain comes from the local references.

We compared full SCALE, which selects K from {4, 8}, with a single-reference control that fixes K = 1. We retained the supervised projection procedure, experts, and gate, refitted SSE in the control’s coordinates, and independently retuned the control under the same 100-trial budget. At 20:1, full SCALE improved balanced accuracy on 18 of 20 datasets, with a mean gain of 0.32 percentage points. See Appendix B.5, Table 10.

W2 / Q3 — Robustness to the choice of minority examples

Thank you for pointing out that our original ten seeds varied model fitting while keeping the retained minority examples fixed. Reviewer c7AP (W3 / Q3) raised the same concern. To address both reviews, we sampled ten independent minority subsets per ratio on the existing cohort of 13 binary diagnostic datasets and four pretrained encoders. Each draw was shared across methods, the test sets stayed fixed, and we repeated training-only tuning with the same budget.

At both 10:1 and 20:1, SCALE achieved higher cohort-average balanced accuracy than cRT and logit adjustment on all ten draws. At 20:1, the paired gains were 4.50 ± 0.36 points over cRT and 2.65 ± 0.42 points over logit adjustment (mean ± SD across the ten cohort-average gains). Appendix H.5, Tables 28 and 30, reports the paired gains, their variability, and the frequency of individual losses at both ratios.

The current gate corrects toward one designated minority class per task. Simultaneous correction toward several rare classes remains future work.

Reviewer_78zMOpenReview ↗

Response to authors

Thanks for running the extra experiments and updating the paper. The standalone RBF-SVM results and the minority resampling tests answer Q1 and Q3 well, and adding the note on multiple rare classes makes the scope clear. On Q2, Table 10 is helpful, though it also shows that K = 1 already gets nearly the same overall accuracy as full SCALE (77.25% vs. 77.38%), with local centers helping a bit more mainly at 20:1. It would be good to mention this overall number directly in Section 4.4 so readers do not overestimate the role of local clustering. Overall, my questions are answered and I will keep my Weak Accept score.

Thank you for the follow-up and for reviewing the additional experiments. We will add the overall comparison to Section 4.4 and clarify the contribution of local references.

The K=1K=1 control retains SCALE’s supervised projection, complementary experts, support expansion, and correction gate. Its strong performance therefore also reflects the effectiveness of these shared components. Within this framework, local references further improve overall BA from 77.25% to 77.38%, with a paired 95% CI of [0.07, 0.20] percentage points. The improvement increases to 0.32 points at 20:1. We will make this distinction explicit in the main text.

Thank you again for reviewing our paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.