Pith. sign in

REVIEW 3 major objections 4 minor 46 references

Mixture Experts with Test-Time Self-Supervised Aggregation for Tabular Imbalanced Regression

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read MATI is claimed to be the first tabular imbalanced-regression method whose accuracy holds under balanced, normal, and inverse test distributions, with a 7.1% average MAE gain.

desk verdict A useful evaluation protocol and a plausible aggregation recipe, undermined by an invalid theorem and missing empirical controls. read the letter →

arxiv 2506.07033 v1 pith:3MMARXNX submitted 2025-06-08 cs.LG

classification cs.LG
keywords TabularregressionImbalancedTest-timetrainingMixtureofexpertsGaussianmodelSelf-supervisedlearningDistributionshift
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

MATI is a method for tabular regression when the training labels are imbalanced, and it is built for the realistic situation where the test distribution is not the balanced one the model was tuned for. The paper claims that by training several region-specific experts from Gaussian components of the label distribution and then reweighting those experts at test time using only unlabeled test features, a model can stay accurate on balanced, training-like, and inverted test distributions alike. On four real-world benchmarks (house prices, abalone age, bike sharing), MATI reports an average 7.1% improvement in mean absolute error over existing imbalanced-regression methods. The reason this matters is that current imbalanced-regression methods usually assume the test labels are balanced, an assumption that often fails in deployment.

What carries the argument

The load-bearing mechanism is the pair consisting of the Region-Aware Mixture Expert and Test-Time Self-Supervised Expert Aggregation. A Gaussian Mixture Model on the labels partitions the training data into regions; a SMOGN-style synthesizer oversamples each region using the component's mean and standard deviation, and each of the $n$ experts is trained on its own synthesized dataset. At test time, VIME perturbs each sample into two views, each expert emits a prediction, and a softmax-normalized weight vector is learned by minimizing the continuous prediction gap $S = \frac{1}{|D_t|}\sum_{X\in D_t}(y_1 - y_2)^2$ over the unlabeled test batch. Theorem 1's identity $S \propto I(\hat{Y}; Y)$ is what licenses the training-free reweighting: reducing the gap between views is claimed to maximize the mutual information between predictions and true test labels.

What would settle it

Take a synthetic regression task with a known mixture of three Gaussian label regions, train MATI, and hold out test sets with three different known label distributions. Compare MATI's learned expert weights against the oracle weights that minimize MAE on the true test labels; if the learned weights are not positively correlated with the oracle weights across the three distributions, the stability-expertise premise is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that the two difficulties of imbalanced tabular regression—rare target regions and unknown test distributions—can be handled in one architecture. The training side fits a Gaussian Mixture Model to the target variable, assigns every training sample to a component, and then oversamples each component's target range with a regression synthesizer so that a separate expert network specializes in each region. The test-time side freezes those experts and learns only their aggregation weights, minimizing the squared difference between predictions on two lightly perturbed views of each unlabeled test sample, a procedure the paper calls Continuous Prediction Gap Minimization. This self-supervised step is justified by a theorem stating that the prediction gap is proportional to the mutual information between the aggregated prediction and the true test label distribution, so giving more weight to experts whose predictions are stable across views is taken to mean giving more weight to experts that actually know the region. The paper further argues that no prior method for imbalanced regression handles all three test distributions they evaluate.

Load-bearing premise

The load-bearing premise is that when a test sample is lightly altered, the experts that predict it most consistently are the ones that actually know its region best.

Editorial extensions

If this is right

  • Existing imbalanced-regression systems can be evaluated under distribution shift instead of only on balanced test sets.
  • A deployed regressor can adapt to a shifted test distribution using only unlabeled features, updating just a handful of expert weights.
  • Rare target regions, such as very expensive houses or extreme bike-rental counts, receive dedicated expert models rather than being absorbed into majority-region predictions.
  • The same evaluation protocol—balanced, normal, and inverse test sets—can serve as a standard stress test for future tabular imbalance methods.
  • If the reported averages hold, MATI is the first tabular imbalanced-regression method that stays competitive when the test distribution is normal, balanced, or the inverse of the training distribution.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural next test is to apply the same self-supervised reweighting to arbitrary regressor ensembles, since nothing in the stability-expertise argument is specific to GMM-trained experts.
  • If the theorem's proportionality holds empirically, the learned weights should track the bin-level density of test labels; one could verify this on synthetic data with known test distributions.
  • The method's reliance on only test features makes it attractive for privacy-sensitive deployments, but it also means the quality of perturbation is a sensitive knob, and the paper already shows performance degrades sharply when the corruption ratio exceeds 0.4.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes MATI, a mixture-of-experts method for tabular imbalanced regression. Training splits the target space via a Gaussian Mixture Model, synthesizes region-specific datasets using a SMOGN-style oversampler, and trains separate TabNet experts on each region. At test time, the experts are frozen and per-expert softmax weights are learned by minimizing the prediction gap between two VIME-perturbed views of each unlabeled test sample. The authors evaluate on four datasets under balanced, normal, and inverse test distributions and report a 7.1% average MAE improvement over existing imbalanced regression methods. A theoretical claim (Theorem 1) states that the prediction gap is positively proportional to mutual information between predictions and test labels.

Significance. The problem addressed is real: existing imbalanced regression methods generally assume a balanced test distribution, and test-time adaptation to unknown test distributions is a worthwhile goal. The proposed evaluation protocol with normal and inverse test sets is a useful contribution, and the region-aware expert training idea is reasonable. The empirical study covers four datasets and multiple distribution shifts. However, the paper's central theoretical justification for test-time aggregation, Theorem 1, is invalid as stated, and there is no ablation that isolates the test-time aggregation mechanism from the region-aware training. These gaps are load-bearing for the novelty claim, so the current version does not yet support the main conclusion.

major comments (3)
  1. [Section 3.3, Eq. (7)] The algebraic chain in Eq. (7) is incorrect. For a bin Z_k with n = |Z_k| elements and center c_k = (1/n)∑ y_j, the identity is ∑_{j,j'}(y_j - y_{j'})^2 = 2n ∑_j (y_j - c_k)^2. Therefore the left-hand side of Eq. (7), ∑_{j,j'} -(y_j - y_{j'})^2, is non-positive, whereas the final expression ∑_j ||y_j - c_k||^2 is non-negative. The first equality in Eq. (7) is also wrong by a factor of 1/n. Consequently, Eq. (7) does not establish a sign-consistent or scale-consistent relationship between the prediction gap and center loss. The subsequent assertion that S ∝ I(Ŷ; Y) is unsupported: the step from binned predictions to mutual information with the true test labels is asserted rather than derived, and S is defined in Eq. (6) as an empirical batch average while Theorem 1 treats it as a population quantity. Since the stability-implies-expertise premise in Section 3.3 is the foundation of test-time aggregation and Theorem 1 is intended to justify it, this is a load-bearing flaw.
  2. [Section 4.3 and Tables 3-4] No ablation isolates the test-time self-supervised aggregation step. Table 4 only reports the learned weights; it does not show what performance would be with uniform weights, with validation-set-fixed weights, or without aggregation (e.g., a single expert or simple average). Without such comparisons, the reported improvements cannot be attributed to the proposed aggregation mechanism rather than to region-aware synthetic training, the new evaluation protocol, or hyperparameter choices. Please add ablations comparing MATI with (i) uniform expert weights, (ii) weights fixed from labeled validation data, and (iii) a single expert trained on the overall synthesized dataset. This is essential because the claimed novelty over existing region-aware training hinges on the aggregation step.
  3. [Section 4.1 and Tables 1-2] The empirical claims lack variance estimates and significance testing. Results are reported as averages over three seeds, but no standard deviations or significance tests are given, and several cells show MATI worse than Vanilla on the normal test distribution (e.g., Table 1 CHW MAPE normal: MATI 22.793 vs. Vanilla 21.714; Table 2 Abalone normal MAE: MATI 2.079 vs. Vanilla 1.947). Given the modest margins on some cells, the 7.1% average improvement claim needs error bars and per-dataset significance information to be credible.
minor comments (4)
  1. [Section 3.2, Eq. (2)] The notation is inconsistent: Eq. (2) uses σ_n for the Gaussian parameter and the text then refers to both σ_n and Σ_n; please use a single consistent symbol, e.g., variance σ_n^2, and distinguish it from the mixture weights π_n.
  2. [Algorithm 1, lines 7-10] Line 10 writes ŷ = wv without defining v; please write the per-expert prediction v_i explicitly and clarify that the softmax is applied over the weight vector w before the weighted sum.
  3. [Section 3.3] The terms 'logits' and 'mutual information' are used loosely in a regression context; consider replacing 'logit output' with 'scalar prediction' and stating precisely which random variables the mutual information I(Ŷ; Y) is computed over.
  4. [Section 4.4 and Figure 4] The perturbation-ratio experiment is summarized only verbally and the figure is not described in text; please state the evaluation metric used, the range of ratios, and the exact protocol for generating the two views.

Circularity Check

1 steps flagged · score 2.0 of 10

Only peripheral circularity: the region-expert validation in Section 4.3 reuses the fitted GMM to build both training and test regions; the central empirical result is not circular.

  1. fitted input called prediction [Section 4.3, Table 3 (with Eq. 5 in Section 3.2)]
    "Region-specific test sets were created by sampling based on Gaussian means and standard deviations, enabling performance evaluation of each expert in its corresponding region. Table 3 confirms that each expert excels within its designated region, validating the effectiveness of Gaussian Mixture Synthesizing in modeling region-specific characteristics."

    The GMM fitted on the training labels defines both the expert-specific oversampling interval in Eq. (5), [mu_n - alpha*sigma_n, mu_n + alpha*sigma_n], and the 'region-specific test sets', which are 'created by sampling based on Gaussian means and standard deviations'. Thus Expert_n is trained on synthetic data drawn from the same region on which it is later evaluated, so its superior performance in 'its' region is forced by the evaluation design rather than by an independent property of the mixture. Presenting this consistency as a validation of the Gaussian Mixture Synthesizing is a fitted input being reported as a prediction. The main results in Tables 1-2 and the weight-shift observation in Table 4 do not share this circularity.

full rationale

The paper's central empirical claim is a 7.1% average MAE improvement over baselines on held-out balanced/normal/inverse test distributions; that comparison is self-contained and does not reduce to a fitted parameter. The theoretical Theorem 1 is not circular in the sense of assuming its conclusion: it attempts a syllogism via center loss, cited as proportional to mutual information in classification [46,41], and Eq. (7). The theorem is not trustworthy, but the flaw is an invalid algebraic identity (a non-positive pairwise-difference sum cannot equal a non-negative center loss) and a conflation of the view-consistency gap of Eq. (6) with bin-wise pairwise label differences. This is a correctness gap rather than a definitional circularity. The only concrete circular-by-construction element is the region-specialization ablation in Section 4.3, where the same GMM parameters define both the training range and the test region. Because that ablation is a secondary sanity check and the authors' own prior-work citations are only for dataset/domain context, the overall circularity score is low.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

MATI introduces no new physical or ontological entities. Its central postulates are modeling assumptions about GMM regions, synthesis quality, perturbation semantics, and the stability-expertise link. The most fragile of these is the stability-expertise premise, which the theorem does not rigorously establish.

free parameters (5)
  • Number of Gaussian components N = 3 (CHW, KHH, Bike Sharing), 2 (Abalone), selected by AIC
    Section 3.2: the number of experts equals the number of GMM components, chosen by Akaike Information Criterion per dataset. Different N changes the expert partition and downstream performance.
  • Oversampling range multiplier alpha = 1 to 2
    Section 3.2, Eq. (5): alpha scales the standard deviation that defines each Gaussian region's target range used for synthesizing expert training data. The paper states alpha is set from 1 to 2 without per-dataset disclosure of the exact value.
  • VIME corruption ratio r = 0.1
    Section 4.1: fixed corruption ratio for test-time perturbations. Figure 4 shows sensitivity at higher ratios, so this value affects the aggregation objective.
  • Test-time adaptation epochs T' = 20 to 50, depending on test set size
    Section 4.1: the number of epochs for test-time weight optimization is chosen based on test set size, introducing a dataset-dependent choice not determined by a stated rule.
  • Target bin resolution delta_y = 1 million (house), 1 (Abalone), 5 (Bike Sharing)
    Section 4.1: bin width affects the LDS weighting baselines, the construction of balanced/normal/inverse test sets, and the evaluation metrics. It is a protocol choice that influences all compared methods.
assumptions (5)
  • domain assumption The GMM fitted on target labels defines meaningful regions that align with challenge regions of the regression function.
    Section 3.2: data are divided by posterior of GMM on Y; if these clusters do not correspond to distinct feature-target regimes, the synthesized per-expert datasets do not create useful specialists.
  • domain assumption SMOGN-based synthesis (Synthesizer and Synthesizer_ex) produces training data whose distribution shift improves expert performance in the corresponding target region.
    Section 3.2 relies on over-sampling target ranges to teach experts; no analysis of synthetic-to-real generalization is provided.
  • ad hoc to paper Prediction stability (small output gap between two VIME perturbations of the same test input) is positively correlated with true expertise on that input.
    Section 3.3 states this as an intuition and Theorem 1 is intended to justify it. The proof is a sketch with an algebraic step in Eq. (7) that is not established, so the premise is effectively assumed.
  • domain assumption Center-loss to mutual-information proportionality, proven for classification, transfers to regression by binning continuous targets.
    Section 3.3 Theoretical Analysis imports results from [46,41]. The binning step assumes prediction bins align with test-label bins, which is not guaranteed.
  • domain assumption VIME perturbations preserve the semantic meaning of a tabular test sample, so output disagreement is a stability signal rather than label noise.
    Used in Algorithm 1; if perturbations alter target-relevant features, gap minimization may adapt to noise rather than to the underlying test distribution.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mixture Experts with Test-Time Self-Supervised Aggregation for Tabular Imbalanced Regression." pith.science (2026). https://pith.science/paper/3MMARXNX

@misc{pith2026250607033,
  author       = {Pith},
  title        = {Pith review of: Mixture Experts with Test-Time Self-Supervised Aggregation for Tabular Imbalanced Regression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3MMARXNX}},
  note         = {Machine review of arXiv:2506.07033}
}
read the original abstract

Tabular data serve as a fundamental and ubiquitous representation of structured information in numerous real-world applications, e.g., finance and urban planning. In the realm of tabular imbalanced applications, data imbalance has been investigated in classification tasks with insufficient instances in certain labels, causing the model's ineffective generalizability. However, the imbalance issue of tabular regression tasks is underexplored, and yet is critical due to unclear boundaries for continuous labels and simplifying assumptions in existing imbalance regression work, which often rely on known and balanced test distributions. Such assumptions may not hold in practice and can lead to performance degradation. To address these issues, we propose MATI: Mixture Experts with Test-Time Self-Supervised Aggregation for Tabular Imbalance Regression, featuring two key innovations: (i) the Region-Aware Mixture Expert, which adopts a Gaussian Mixture Model to capture the underlying related regions. The statistical information of each Gaussian component is then used to synthesize and train region-specific experts to capture the unique characteristics of their respective regions. (ii) Test-Time Self-Supervised Expert Aggregation, which dynamically adjusts region expert weights based on test data features to reinforce expert adaptation across varying test distributions. We evaluated MATI on four real-world tabular imbalance regression datasets, including house pricing, bike sharing, and age prediction. To reflect realistic deployment scenarios, we adopted three types of test distributions: a balanced distribution with uniform target frequencies, a normal distribution that follows the training data, and an inverse distribution that emphasizes rare target regions. On average across these three test distributions, MATI achieved a 7.1% improvement in MAE compared to existing methods.

Figures

Figures reproduced from arXiv: 2506.07033 by the authors.

Figure 1
Figure 1. Illustration of MATI. (a) Existing works on imbalanced regression focus on [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Method overview of MATI, where DT denotes the imbalanced training data, DS denotes the dataset after overall synthesizing. Dn S denotes the datasets after the second synthesizing that considered statistical information from each Gaussian component. The expert training part in purple trains region-aware expert models. The expert aggregation part in yellow aggregates the expert models with unlabeled test data Dt. 16, … view at source ↗
Figure 2
Figure 2. To mitigate the imbalance by generating synthetic samples in under [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figures from the paper (2 more)
Figure 3
Figure 3. Figure 3: Distribution of the four DIR datasets. Datasets. We curated four real-world tabular imbalance regression benchmarks across diverse domains, as shown in 3, including pricing, aging, and counts, and constructed three distinct test data distributions: balanced, normal, an…
Figure 4
Figure 4. Figure 4: The relationship between the performance of the four datasets and the [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 32 canonical work pages

  1. [1]

    Electronics11(4), 662 (2022)

    Alfaiz, N.S., Fati, S.M.: Enhanced credit card fraud detection model using machine learning. Electronics11(4), 662 (2022)

  2. [2]

    In: Pro- ceedings of the AAAI conference on artificial intelligence

    Arik, S.Ö., Pfister, T.: Tabnet: Attentive interpretable tabular learning. In: Pro- ceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 6679–6687 (2021)

  3. [3]

    In: First international workshop on learning with imbalanced domains: Theory and applications

    Branco, P., Torgo, L., Ribeiro, R.P.: Smogn: a pre-processing approach for im- balanced regression. In: First international workshop on learning with imbalanced domains: Theory and applications. pp. 36–50. PMLR (2017)

  4. [4]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Cai, J., Wang, Y., Hwang, J.N.: Ace: Ally complementary experts for solving long- tailed recognition in one-shot. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 112–121 (October 2021)

  5. [5]

    Advances in neural information process- ing systems32(2019)

    Cao, K., Wei, C., Gaidon, A., Arechiga, N., Ma, T.: Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information process- ing systems32(2019)

  6. [6]

    Journal of artificial intelligence research16, 321– 357 (2002)

    Chawla, N.V., Bowyer, K.W., Hall, L.O., Kegelmeyer, W.P.: Smote: synthetic mi- nority over-sampling technique. Journal of artificial intelligence research16, 321– 357 (2002)

  7. [7]

    Knowledge-Based Systems252, 109393 (2022)

    Chen,Z.,Chen,J.,Xie,Z.,Xu,E.,Feng,Y.,Liu,S.:Multi-expertattentionnetwork with unsupervised aggregation for long-tailed fault diagnosis under speed variation. Knowledge-Based Systems252, 109393 (2022)

  8. [8]

    In: Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXIX 16

    Chu, P., Bian, X., Liu, S., Ling, H.: Feature space augmentation for long-tailed data. In: Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXIX 16. pp. 694–710. Springer (2020)

Show all 46 references
  1. [9]

    In: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition

    Cui, Y., Jia, M., Lin, T.Y., Song, Y., Belongie, S.: Class-balanced loss based on effective number of samples. In: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition. pp. 9268–9277 (2019)

  2. [10]

    IEEE transactions on pattern analysis and machine intelligence 41(6), 1367–1381 (2018)

    Dong, Q., Gong, S., Zhu, X.: Imbalanced deep learning by minority class incremen- tal rectification. IEEE transactions on pattern analysis and machine intelligence 41(6), 1367–1381 (2018)

  3. [11]

    In: CIKM

    Du, W., Wang, W., Peng, W.: Dora: Domain-based self-supervised learning frame- work for low-resource real estate appraisal. In: CIKM. pp. 4552–4558. ACM (2023)

  4. [12]

    UCI Machine Learning Repository (2013), DOI: https://doi.org/10.24432/C5W894

    Fanaee-T, H.: Bike Sharing. UCI Machine Learning Repository (2013), DOI: https://doi.org/10.24432/C5W894

  5. [13]

    Advances in Neural Information Processing Systems36(2024) Test-Time Aggregated Experts for Tabular Imbalanced Regression 15

    Gao, J., Zhao, H., Li, Z., Guo, D.: Enhancing minority classes by mixing: an adaptative optimal transport approach for long-tailed classification. Advances in Neural Information Processing Systems36(2024) Test-Time Aggregated Experts for Tabular Imbalanced Regression 15

  6. [14]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

    Guo,H.,Wang,S.:Long-tailedmulti-labelvisualrecognitionbycollaborativetrain- ing on uniform and re-balanced samplings. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. pp. 15089–15098 (2021)

  7. [15]

    Guo, T., Chen, X., Wang, Y., Chang, R., Pei, S., Chawla, N.V., Wiest, O., Zhang, X.: Large language model based multi-agents: A survey of progress and challenges (2024), https://arxiv.org/abs/2402.01680

  8. [16]

    He,H.,Garcia,E.A.:Learningfromimbalanceddata.IEEETransactionsonknowl- edge and data engineering21(9), 1263–1284 (2009)

  9. [17]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Hong, Y., Han, S., Choi, K., Seo, S., Kim, B., Chang, B.: Disentangling label distribution for long-tailed visual recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6626–6636 (2021)

  10. [18]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Huang, C., Li, Y., Loy, C.C., Tang, X.: Learning deep representation for imbal- anced classification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5375–5384 (2016)

  11. [19]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Jamal, M.A., Brown, M., Yang, M.H., Wang, L., Gong, B.: Rethinking class- balanced methods for long-tailed visual recognition from a domain adaptation perspective. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7610–7619 (2020)

  12. [20]

    Jiang, A.Q., Sablayrolles, A., Roux, A., Mensch, A., Savary, B., Bamford, C., Chaplot, D.S., de las Casas, D., Hanna, E.B., Bressand, F., Lengyel, G., Bour, G., Lample, G., Lavaud, L.R., Saulnier, L., Lachaux, M.A., Stock, P., Subramanian, S., Yang, S., Antoniak, S., Scao, T.L...

  13. [21]

    arXiv preprint arXiv:1910.09217 (2019)

    Kang, B., Xie, S., Rohrbach, M., Yan, Z., Gordo, A., Feng, J., Kalantidis, Y.: Decoupling representation and classifier for long-tailed recognition. arXiv preprint arXiv:1910.09217 (2019)

  14. [22]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Kim, J., Jeong, J., Shin, J.: M2m: Imbalanced classification via major-to-minor translation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13896–13905 (2020)

  15. [23]

    Sociological methods & research33(2), 188–229 (2004)

    Kuha, J.: Aic and bic: Comparisons of assumptions and performance. Sociological methods & research33(2), 188–229 (2004)

  16. [24]

    International Jour- nal of Data Science and Analytics12(2), 165–174 (2021)

    Lebichot, B., Paldino, G.M., Siblini, W., He-Guelton, L., Oblé, F., Bontempi, G.: Incremental learning strategies for credit cards fraud detection. International Jour- nal of Data Science and Analytics12(2), 165–174 (2021)

  17. [25]

    In: PKDD (4)

    Li,C.,Wang,W., Du,W., Peng, W.:Lookaround! Aneighbor relationgraph learn- ing framework for real estate appraisal. In: PKDD (4). Lecture Notes in Computer Science, vol. 14648, pp. 3–16. Springer (2024)

  18. [26]

    In: Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition

    Li, M., Cheung, Y.m., Lu, Y.: Long-tailed visual recognition via gaussian clouded logit adjustment. In: Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition. pp. 6929–6938 (2022)

  19. [27]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Li, M., Zhikai, H., Lu, Y., Lan, W., Cheung, Y.m., Huang, H.: Feature fusion from head to tail for long-tailed visual recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 13581–13589 (2024)

  20. [28]

    arXiv preprint arXiv:2007.07314 (2020)

    Menon, A.K., Jayasumana, S., Rawat, A.S., Jain, H., Veit, A., Kumar, S.: Long-tail learning via logit adjustment. arXiv preprint arXiv:2007.07314 (2020)

  21. [29]

    Nash, W., Sellers, T., Talbot, S., Cawthorn, A., Ford, W.: Uci machine learning repository: Abalone data set (1994), https://archive.ics.uci.edu/ml/datasets/abalone, accessed: 2024-07-30

  22. [30]

    In: Proceedings 16 Y.-C

    Park,S.,Hong,Y.,Heo,B.,Yun,S.,Choi,J.Y.:Themajoritycanhelptheminority: Context-rich minority oversampling for long-tailed classification. In: Proceedings 16 Y.-C. Wang et al. of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6887–6896 (2022)

  23. [31]

    In: International conference on learning representations (2021)

    Peng, H., Sun, M., Li, P.: Optimal transport for long-tailed recognition with learn- able cost matrix. In: International conference on learning representations (2021)

  24. [32]

    Advances in neural information processing systems33, 4175–4186 (2020)

    Ren, J., Yu, C., Ma, X., Zhao, H., Yi, S., et al.: Balanced meta-softmax for long- tailed visual recognition. Advances in neural information processing systems33, 4175–4186 (2020)

  25. [33]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Ren, J., Zhang, M., Yu, C., Liu, Z.: Balanced mse for imbalanced visual regression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7926–7935 (2022)

  26. [34]

    Machine Learning109, 1803–1835 (2020)

    Ribeiro, R.P., Moniz, N.: Imbalanced regression and extreme value prediction. Machine Learning109, 1803–1835 (2020)

  27. [35]

    Machine Learning110, 2187–2211 (2021)

    Steininger, M., Kobs, K., Davidson, P., Krause, A., Hotho, A.: Density-based weighting for imbalanced regression. Machine Learning110, 2187–2211 (2021)

  28. [36]

    Advances in neural information processing systems33, 8101– 8113 (2020)

    Tian, J., Liu, Y.C., Glaser, N., Hsu, Y.C., Kira, Z.: Posterior re-calibration for imbalanced datasets. Advances in neural information processing systems33, 8101– 8113 (2020)

  29. [37]

    In: Por- tuguese conference on artificial intelligence

    Torgo, L., Ribeiro, R.P., Pfahringer, B., Branco, P.: Smote for regression. In: Por- tuguese conference on artificial intelligence. pp. 378–389. Springer (2013)

  30. [38]

    ACM Transactions on Spatial Algorithms and Systems9(1), 1–24 (2023)

    Wang,D.,Fu,Y.,Liu,K.,Chen,F.,Wang,P.,Lu,C.T.:Automatedurbanplanning for reimagining city configuration via adversarial learning: quantification, genera- tion, and evaluation. ACM Transactions on Spatial Algorithms and Systems9(1), 1–24 (2023)

  31. [39]

    arXiv preprint arXiv:2010.01809 (2020)

    Wang, X., Lian, L., Miao, Z., Liu, Z., Yu, S.X.: Long-tailed recognition by routing diverse distribution-aware experts. arXiv preprint arXiv:2010.01809 (2020)

  32. [40]

    Advances in Neural Information Processing Systems36(2024)

    Wang, Z., Wang, H.: Variational imbalanced regression: Fair uncertainty quan- tification via probabilistic smoothing. Advances in Neural Information Processing Systems36(2024)

  33. [41]

    In: Computer vision–ECCV 2016: 14th European conference, amsterdam, the netherlands, October 11–14, 2016, proceedings, part VII 14

    Wen, Y., Zhang, K., Li, Z., Qiao, Y.: A discriminative feature learning approach for deep face recognition. In: Computer vision–ECCV 2016: 14th European conference, amsterdam, the netherlands, October 11–14, 2016, proceedings, part VII 14. pp. 499–515. Springer (2016)

  34. [42]

    In: International conference on machine learning

    Yang, Y., Zha, K., Chen, Y., Wang, H., Katabi, D.: Delving into deep imbal- anced regression. In: International conference on machine learning. pp. 11842– 11851. PMLR (2021)

  35. [43]

    Yigitcanlar, T., Kankanamge, N., Regona, M., Ruiz Maldonado, A., Rowan, B., Ryu, A., Desouza, K.C., Corchado, J.M., Mehmood, R., Li, R.Y.M.: Artificial intelligence technologies and related urban planning and development concepts: How are they perceived and utilized in austral...

  36. [44]

    Advances in Neural Information Processing Systems33, 11033–11043 (2020)

    Yoon, J., Zhang, Y., Jordon, J., Van der Schaar, M.: Vime: Extending the suc- cess of self-and semi-supervised learning to tabular domain. Advances in Neural Information Processing Systems33, 11033–11043 (2020)

  37. [45]

    arXiv preprint arXiv:2301.08915 (2023)

    Zhang, S., Yang, L., Mi, M.B., Zheng, X., Yao, A.: Improving deep regression with ordinal entropy. arXiv preprint arXiv:2301.08915 (2023)

  38. [46]

    Advances in Neural Information Processing Systems35, 34077–34090 (2022)

    Zhang, Y., Hooi, B., Hong, L., Feng, J.: Self-supervised aggregation of diverse experts for test-agnostic long-tailed recognition. Advances in Neural Information Processing Systems35, 34077–34090 (2022)

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.