acceptodds
Under review as a conference paper at ICLR 2027

DIDA: A Value-Driven Paradigm for Data Importance-Based Tabular Data Augmentation in Low-Data High-Constraint Scenarios

Abstract

In regulated domains such as medical diagnosis and financial risk control, tabular datasets are often small, imbalanced, and subject to strong cross-feature constraints. Existing augmentation methods therefore face a difficult trade-off among preserving semantic consistency, increasing sample diversity, and keeping augmentation decisions auditable. We propose DIDA (Data-Importance driven Differential Augmentation), a value-aware framework that differentiates augmentation at the sample-feature level using two complementary signals: feature-level SHAP importance and sample-level Shapley contribution. We formulate augmentation as diversity maximization under semantic-consistency constraints and show that the optimal perturbation magnitude decreases with sample contribution and scales inversely with feature importance (Theorem 1), providing a principled basis for value-dependent perturbation. DIDA further incorporates dynamic domain correction and constraint-guided categorical augmentation to preserve valid feature combinations. Across six public tabular datasets, six augmentation baselines, and ten downstream classifiers, DIDA improves average AUC-ROC, AUC-PR, and F1 by 1.94, 2.30, and 4.30 percentage points, respectively, over the strongest competing baseline. Its generated samples also achieve lower JS divergence and Wasserstein distance, together with higher information entropy and feature coverage, indicating that DIDA improves diversity without sacrificing distributional fidelity.

Primary area:
datasets and benchmarks
Keywords:
Data-Centric AI, Tabular Data Augmentation, Data Importance, Class Imbalance

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_PQ1VOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DIDA, a value driven tabular data augmentation framework for low data, high constraint scenarios. DIDA uses SHAP to estimate feature level importance and KNN Shapley to estimate sample level contribution, and uses these two signals to guide enhancement of sample feature levels. This method also introduces dynamic domain correction to preserve high-value boundary samples, constructs constraint tables from cross feature associations to guide classification enhancement, and applies value related perturbations to numerical features.

Critical strengths

1.DIDA targets a practically important low data, high constraint setting, in which traditional enhancement methods are prone to violate feature dependency relationships or distort information samples. Therefore, its value consciousness and constraint oriented design are very suitable for such applications.

2.The proposed enhancement mechanism is relatively interpretable and auditable. DIDA utilizes interpretable methods such as feature level importance and sample level contribution for generation, allowing enhanced decisions to be traced back to identifiable value scores and constraint rules, rather than opaque generation processes. This is particularly useful in regulated or high-risk spreadsheet applications, where understanding the reasons for sample modifications is equally important as predictive performance.

Critical weaknesses

1.DIDA uses high contribution samples to expand various feature domains SjS_j. If a high-value sample violates substantive business rules rather than just statistical boundaries, this correction step may make previously invalid values "valid" and be included in the category of data augmentation. Furthermore, even if the range of values for a single feature is met, it may not necessarily ensure that the generated combination of numerical features conforms to the real-world business constraints across features. If the difference between "statistical outliers" and "violation of strict business rules" can be more clearly defined, and constraint verification can be clearly performed after correction, it will help improve the security and applicability of this method in strong constraint application scenarios.

2.When the precise constraint table lookup fails, DIDA gradually relaxes the conditional constraints, and it is unclear whether the final complete feature combination still meets the original constraints. The current ablation mainly evaluates downstream predictive performance, rather than directly reporting constraint violation rates (CVR) at different rollback levels. Direct evaluation of invalid or unsupported combinations, especially after relaxation, margin, and uniform rollback, is crucial for verifying the claimed business compliance assurance.

Questions for authors

See Weaknesses

6 comments

Thank you for carefully reviewing our paper and recognizing the practical significance of small-sample settings and settings with strong constraints, as well as the interpretability and auditability of value-aware augmentation. Your questions about statistical boundaries, hard business rules, and the validity of combinations after fallback have also prompted us to further consider the relationships among these design choices; below, we explain them in terms of our design considerations and supplementary results.

W1 Dynamic Domain Correction, Hard Business Rules, and Cross-Feature Constraints

In response to your comments on domain correction and business constraints, we discuss three aspects: the considerations behind introducing this mechanism, issues that currently require further study, and experiments in small-sample settings.

(1) We introduced domain correction mainly because statistical boundaries in small-sample settings may be too narrow. The range obtained using the interquartile range reflects the region where the current data are concentrated, and some genuine, rare observations at the margins that carry important predictive information may fall outside this range. If these observations are directly excluded, subsequent augmentation can only take place within a further narrowed range, and the already limited information at the margins can easily be lost.

(2) Regarding your suggestion to distinguish statistical outliers from violations of business rules and to validate constraints after correction, we believe that more detailed mathematical modeling is still needed to clarify the relationships among statistical boundaries, sample value, and business constraints. The difficulty is that adjusting the range of a single feature may also affect the validity of its combinations with other features, while the lack of complete domain-rule descriptions in public datasets makes both modeling and validation more difficult. In future work, we will continue to study these relationships and improve joint constraint checks after domain correction and sample generation, to enhance the safety and applicability of the method in settings with strong constraints.

(3) At this stage, we use supplementary experiments at different training fractions to evaluate DIDA's predictive performance in small-sample settings more fully, with the results shown in the table below. DIDA achieves the highest mean ROC-AUC for most combinations of training size and dataset, and also maintains its lead at the smallest training fraction on all six datasets. This shows that the complete method still yields predictive benefits when training information is limited, supporting the practical significance of our design of DIDA for small-sample settings.

Table: ROC-AUC of each method at different training fractions (mean ± sample standard deviation)

DatasetTrain (%)OriginalSMOTECTGANSMOTE-CDNNTabDDPMMDLM-localTANDEMDIDA
Alzheimers25%<u>0.8224 ± 0.0075</u>0.8183 ± 0.00880.7870 ± 0.01230.8171 ± 0.00380.7901 ± 0.00970.8077 ± 0.01330.8107 ± 0.00880.8429 ± 0.0094
50%0.8507 ± 0.00960.8517 ± 0.01210.8314 ± 0.0075<u>0.8535 ± 0.0091</u>0.8301 ± 0.01080.8377 ± 0.00840.8378 ± 0.00590.8654 ± 0.0080
75%<u>0.8630 ± 0.0078</u>0.8617 ± 0.00850.8497 ± 0.01100.8606 ± 0.00990.8447 ± 0.00820.8514 ± 0.00800.8529 ± 0.00690.8708 ± 0.0069
100%<u>0.8672 ± 0.0076</u>0.8654 ± 0.01040.8538 ± 0.00680.8656 ± 0.00750.8508 ± 0.01010.8580 ± 0.00840.8577 ± 0.00910.8745 ± 0.0062
Kidney25%0.6332 ± 0.03130.6398 ± 0.02950.6150 ± 0.03850.6434 ± 0.03150.5933 ± 0.03410.6434 ± 0.0247<u>0.6457 ± 0.0348</u>0.6639 ± 0.0418
50%0.6765 ± 0.03110.6776 ± 0.02710.6495 ± 0.0347<u>0.6837 ± 0.0280</u>0.6244 ± 0.03410.6720 ± 0.02850.6741 ± 0.02050.6945 ± 0.0246
75%0.6939 ± 0.0252<u>0.6998 ± 0.0299</u>0.6609 ± 0.03440.6974 ± 0.03000.6471 ± 0.03420.6930 ± 0.02910.6878 ± 0.04020.7099 ± 0.0281
100%0.7050 ± 0.01670.7100 ± 0.02570.6770 ± 0.02420.7149 ± 0.01900.6589 ± 0.03050.7129 ± 0.03380.6939 ± 0.0172<u>0.7136 ± 0.0187</u>
Heart25%0.8203 ± 0.01530.8127 ± 0.00990.8131 ± 0.02100.8157 ± 0.01470.8261 ± 0.01830.8201 ± 0.0141<u>0.8270 ± 0.0196</u>0.8306 ± 0.0170
50%0.8392 ± 0.01710.8262 ± 0.02260.8277 ± 0.01870.8324 ± 0.0227<u>0.8398 ± 0.0161</u>0.8368 ± 0.02190.8328 ± 0.02010.8489 ± 0.0236
75%0.8453 ± 0.02020.8397 ± 0.02340.8439 ± 0.02670.8393 ± 0.0206<u>0.8472 ± 0.0229</u>0.8414 ± 0.01250.8405 ± 0.02060.8574 ± 0.0247
100%0.8419 ± 0.01290.8416 ± 0.0200<u>0.8621 ± 0.0257</u>0.8442 ± 0.01850.8605 ± 0.03030.8521 ± 0.03430.8518 ± 0.02050.8725 ± 0.0294
Healthcare25%<u>0.7410 ± 0.0101</u>0.7403 ± 0.00700.7402 ± 0.00780.7385 ± 0.00730.7397 ± 0.00480.7285 ± 0.00680.7311 ± 0.00710.7469 ± 0.0052
50%<u>0.7514 ± 0.0065</u>0.7478 ± 0.00680.7489 ± 0.00670.7460 ± 0.00750.7487 ± 0.00640.7387 ± 0.00680.7393 ± 0.00800.7548 ± 0.0059
75%<u>0.7539 ± 0.0061</u>0.7519 ± 0.00560.7516 ± 0.00570.7500 ± 0.00530.7523 ± 0.00510.7441 ± 0.00510.7409 ± 0.00720.7573 ± 0.0045
100%<u>0.7554 ± 0.0058</u>0.7526 ± 0.00490.7529 ± 0.00630.7514 ± 0.00400.7537 ± 0.00440.7471 ± 0.00520.7422 ± 0.00690.7585 ± 0.0050
Teen25%0.8237 ± 0.04820.8369 ± 0.04040.8267 ± 0.04090.8453 ± 0.04910.7036 ± 0.06030.8474 ± 0.0336<u>0.8521 ± 0.0563</u>0.8554 ± 0.0360
50%0.9025 ± 0.0469<u>0.9213 ± 0.0467</u>0.8876 ± 0.03730.9403 ± 0.03920.7223 ± 0.04770.8918 ± 0.04750.9054 ± 0.02860.9151 ± 0.0400
75%0.9143 ± 0.03770.9318 ± 0.03850.9350 ± 0.0212<u>0.9373 ± 0.0249</u>0.7867 ± 0.06410.9251 ± 0.03000.9286 ± 0.03260.9511 ± 0.0268
100%0.9398 ± 0.02030.9514 ± 0.02670.9285 ± 0.0281<u>0.9589 ± 0.0180</u>0.8071 ± 0.04440.9413 ± 0.02170.9458 ± 0.03610.9610 ± 0.0303
Urban25%<u>0.5926 ± 0.0065</u>0.5832 ± 0.01050.5856 ± 0.00820.5812 ± 0.01050.5790 ± 0.00780.5794 ± 0.00940.5777 ± 0.00920.5964 ± 0.0045
50%<u>0.6069 ± 0.0037</u>0.6030 ± 0.00530.5988 ± 0.00590.6011 ± 0.00750.6008 ± 0.00840.6021 ± 0.00650.5912 ± 0.00450.6113 ± 0.0054
75%<u>0.6201 ± 0.0031</u>0.6089 ± 0.00510.6091 ± 0.00400.6081 ± 0.00410.6072 ± 0.00250.6094 ± 0.00520.6029 ± 0.00560.6250 ± 0.0034
100%<u>0.6240 ± 0.0036</u>0.6131 ± 0.00550.6112 ± 0.00490.6151 ± 0.00660.6144 ± 0.00660.6167 ± 0.00400.6110 ± 0.00700.6277 ± 0.0041

W2 Value Protection, Fallback Sampling, and Combination Validity

Following your suggestion for direct evaluation, we conducted two types of checks on the final generated samples; below, we describe violations of domain rules and changes in statistical structure, respectively.

(1) Based on the data content and existing domain knowledge, we identified domain rules that could be explicitly checked and checked the final samples generated by the complete pipeline. Because these public datasets do not provide accurate domain-rule tables, we have not yet identified rules suitable for accurate evaluation for some datasets, and we mark these cases as N/A in the table. DIDA reports zero violation rates for all assessed rules and is the only augmentation method in the table to achieve this across all of them, indicating that the final generated samples perform relatively well under these rules.

Table: Results of checks against verifiable domain rules

DatasetRuleOriginalSMOTECTGANSMOTE-CDNNTabDDPMMDLM-localTANDEMDIDA
AlzheimersBP order8.3786 ± 0.38016.3510 ± 0.900011.6725 ± 9.00095.7808 ± 0.76660.0000 ± 0.00000.0000 ± 0.00009.4710 ± 4.58620.0000 ± 0.0000
KidneyBP order7.7961 ± 0.76955.6445 ± 1.05817.8612 ± 4.75135.9319 ± 1.02170.0000 ± 0.00000.0000 ± 0.00005.3982 ± 2.62960.0000 ± 0.0000
HeartNot assessedN/AN/AN/AN/AN/AN/AN/AN/A
HealthcareBMI identity0.0000 ± 0.00001.9397 ± 0.256998.0215 ± 0.39802.0820 ± 0.215090.2260 ± 5.09257.7325 ± 0.919996.0295 ± 0.59290.0000 ± 0.0000
BP order0.0000 ± 0.00000.0000 ± 0.00000.0023 ± 0.00610.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.00000.1241 ± 0.22780.0000 ± 0.0000
TeenNot assessedN/AN/AN/AN/AN/AN/AN/AN/A
UrbanNot assessedN/AN/AN/AN/AN/AN/AN/AN/A
Reviewer_jjVvOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DIDA, a value-aware tabular data augmentation framework for low-data, high-constraint classification settings. It combines feature-level SHAP importance and sample-level Shapley contributions to control perturbations at the sample-feature level. The method includes adaptive numerical perturbation, dynamic domain correction, and constraint-guided categorical augmentation based on estimated feature dependencies.

The authors further formulate augmentation as a constrained diversity-maximization problem and derive a relationship between data importance and perturbation magnitude using KKT conditions. Experiments on six tabular datasets with ten downstream classifiers report improvements over six augmentation baselines in AUC-ROC, AUC-PR, and F1, along with improvements in distributional fidelity and diversity metrics.

Critical strengths

1. A meaningful problem setting. The paper addresses the challenge of augmenting small tabular datasets while preserving important predictive information and respecting feature dependencies. Combining sample-level and feature-level importance provides an intuitive mechanism for allocating augmentation intensity.

2. Evaluation across diverse augmentation methods and predictors. The authors compare DIDA with interpolation-based and generative augmentation methods across ten downstream classifiers. The experiments also include distributional fidelity, diversity, and component ablations, providing evidence that the method can improve downstream performance under the evaluated conditions.

Critical weaknesses

1. Inconsistency in the theoretical derivation and its connection to the implemented method.

Theorem 1 claims that the optimal perturbation magnitude scales proportionally to sqrt((1 - ds'_i) / b'_j). However, Equation (31) in Appendix D gives a closed-form expression whose dependence on feature importance is proportional to 1 / b'_j rather than 1 / sqrt(b'_j), assuming the remaining terms are fixed for a given sample. This apparent inconsistency directly affects a central theoretical contribution.

Furthermore, the practical perturbation function in Equation (4) is a separately parameterized monotonic function rather than the derived closed-form optimum. The theoretical justification for this particular implementation is therefore unclear.

2. Limited novelty relative to importance-guided and dependency-preserving augmentation.

The central principle of preserving important features while perturbing less important ones has been explored in prior work. For example, AGATa (NeurIPS 2024 TRL Workshop) uses attention-guided tabular augmentation to preserve informative feature relationships, while TabCutMixPlus (ICML 2025) groups correlated features to maintain feature coherence during augmentation.

Although DIDA additionally combines feature-level SHAP and sample-level Shapley contributions, its relationship to these approaches is not sufficiently discussed. The paper would benefit from a clearer explanation of what is methodologically new beyond combining existing importance estimation and constrained augmentation techniques.

3. Insufficient justification of semantic constraint preservation.

DIDA constructs categorical constraint tables from feature associations estimated using ANOVA and mutual information. However, statistical dependencies do not necessarily correspond to valid domain constraints. In particular, the fallback procedure relaxes conditioning information and can eventually sample from marginal or legal feature domains, which does not inherently guarantee the validity of cross-feature combinations.

Therefore, the claim that generated samples satisfy semantic or business constraints requires stronger justification, particularly when explicit domain rules are incomplete or unavailable. The evaluation would also benefit from measuring actual constraint-violation rates rather than relying primarily on distributional similarity and downstream predictive performance.

4. Questionable validity of importance-inverse perturbation.

DIDA assumes that less important samples and features can safely receive stronger perturbations. However, low estimated importance does not guarantee label preservation. Large perturbations may destroy meaningful feature interactions or introduce mislabeled synthetic samples.

Although DIDA restricts perturbation ranges, its theoretical objective maximizes diversity rather than label consistency. A comparison of inverse, uniform, and importance-proportional perturbation would better justify this central design choice.

Questions for authors

Q1. Theoretical correctness. Could the authors clarify the discrepancy between Theorem 1 and Equations (31)–(32) regarding the dependence of optimal perturbation magnitude on feature importance? Additionally, how does the practical perturbation function in Equation (4) relate to the derived optimum?

Q2. Novelty and related work. How does DIDA differ from existing importance-aware augmentation approaches such as AGATa and dependency-preserving methods such as TabCutMixPlus? Can the authors clarify the contribution of joint sample-feature importance beyond these existing principles?

Q3. Constraint validity. How does the conditional-probability constraint table guarantee valid cross-feature combinations, particularly when fallback sampling relaxes conditioning information? Could the authors report explicit domain-constraint violation rates?

Q4. Reliability of data valuation. Since SHAP and KNN-Shapley estimates are computed from limited training data, how sensitive is DIDA to estimation errors or different base predictors? Would the improvements persist with simpler or randomly assigned importance scores?

11 comments

W1 / Q1 Consistency of the Theoretical Derivation and the Practical Perturbation Function

Your question about consistency between the theory and the practical method concerns three points: the scope of the square root in the theorem, the relationship between the full solution and the theorem, and how Equation (4) is designed based on the theory. We address them in turn.

(1) The original expression in Theorem 1 is δij∗∝1−dsi′bj′\delta^*_{ij}\propto\frac{\sqrt{1-ds'_i}}{b'_j}, where the denominator bj′b'_j is outside the square root, and the text also explicitly states that the perturbation magnitude is inversely proportional to bj′b'_j. Therefore, the expression (1−dsi′)/bj′\sqrt{(1-ds'_i)/b'_j} used in the review differs from the original expression in the theorem.

(2) The full solution in Equation (31) is δij∗=1bj′κi(1−dsi′)∑k1/bk′\delta^*_{ij}=\frac{1}{b'_j}\frac{\sqrt{\kappa_i(1-ds'_i)}}{\sqrt{\sum_k1/b'_k}}, where κi=2ε0∑kbk′xik2\kappa_i=2\varepsilon_0\sum_k b'_kx_{ik}^{2}. For the same sample, once the full set of feature weights is fixed, all parts of Equation (31) except 1/bj′1/b'_j are the same across features. Thus, the perturbation magnitude for each feature is inversely proportional to bj′b'_j, consistent with Theorem 1 and Equation (32).

(3) The proof of Theorem 1 is intended to analyze how data value affects perturbation allocation under a specified consistency requirement. We designed Equation (4) based on the resulting relationship so that it retains the principle that perturbation intensity does not increase as value increases, while keeping the intensity within a bounded range and avoiding excessively large perturbations when feature importance approaches zero. This also makes it easier to adjust the parameters to the characteristics of different datasets, allowing the theoretical analysis to better inform practical augmentation.

W2 / Q2 Novelty, Related Work, and Joint Guidance by Sample and Feature Value

Regarding the differences between DIDA and existing methods, we would like to explain our contributions from three aspects: the use of the two types of value, the mathematical modeling of the problem, and the design of the entire augmentation pipeline.

(1) AGATa selects low-importance features for modification based on within-sample attention, whereas TabCutMixPlus preserves dependencies by swapping groups of correlated features. DIDA additionally estimates each sample's contribution to the learning task separately and combines it with feature importance, considering both which samples are more valuable and which features within those samples should be perturbed more cautiously.

(2) We consider DIDA's main contribution to be formalizing how the two types of value jointly affect augmentation. We formulate a diversity-maximization model under a specified consistency requirement and, under the corresponding quadratic approximation, use the KKT conditions to derive the relationship between value and perturbation magnitude. This provides a mathematical basis under the corresponding conditions for why perturbations should be allocated according to value and why important information should be modified more cautiously, and also provides guidance for the design of the practical method.

(3) We also incorporate sample value and feature value into the design of the entire augmentation pipeline, allowing them to jointly affect augmentation allocation and generation control. In this way, both types of value are used throughout the pipeline, and the allocation relationship obtained in the theory can inform the design of the practical method. This is the further contribution we aim to make beyond existing importance-guided and dependency-preserving methods.

W3 / Q3 Rule Inputs, Conditional Tables, and Combination Validity after Fallback

To address your question about the conditional table and combination validity after fallback, we combine domain-rule checks and statistical results and answer from the following two aspects.

(1) Because these public datasets do not provide precise domain-rule tables, we identified applicable rules using the data and existing domain knowledge, checked the final generated samples against them, and report the violation rates in the table below. N/A indicates that no domain rule suitable for accurate evaluation has yet been identified, so the corresponding check was not performed. According to the results, DIDA is the only augmentation method in the table to report zero violations across all assessed rules, indicating that it also performs well on the cross-feature rules checked here.

Table: Checks against Verifiable Domain Rules

DatasetRuleOriginalSMOTECTGANSMOTE-CDNNTabDDPMMDLM-localTANDEMDIDA
AlzheimersBP order8.3786 ± 0.38016.3510 ± 0.900011.6725 ± 9.00095.7808 ± 0.76660.0000 ± 0.00000.0000 ± 0.00009.4710 ± 4.58620.0000 ± 0.0000
KidneyBP order7.7961 ± 0.76955.6445 ± 1.05817.8612 ± 4.75135.9319 ± 1.02170.0000 ± 0.00000.0000 ± 0.00005.3982 ± 2.62960.0000 ± 0.0000
HeartNot assessedN/AN/AN/AN/AN/AN/AN/AN/A
HealthcareBMI identity0.0000 ± 0.00001.9397 ± 0.256998.0215 ± 0.39802.0820 ± 0.215090.2260 ± 5.09257.7325 ± 0.919996.0295 ± 0.59290.0000 ± 0.0000
BP order0.0000 ± 0.00000.0000 ± 0.00000.0023 ± 0.00610.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.00000.1241 ± 0.22780.0000 ± 0.0000
TeenNot assessedN/AN/AN/AN/AN/AN/AN/AN/A
UrbanNot assessedN/AN/AN/AN/AN/AN/AN/AN/A

(2) For the statistical assessment, we examine how the value ranges and feature combinations of generated samples differ from the reference data and whether the results are stable across runs. Numeric range is the proportion of numeric cells outside the statistical bounds; Joint support is the proportion of discretized feature-group combinations not observed in the filtered training reference support set. DIDA has relatively low numeric-range deviation, and both the mean and standard deviation of Joint support are lower than those of the other augmentation methods on all six datasets. This shows that our method generates fewer combinations absent from the training reference support set, with less variation in this metric across runs. Together with the preceding rule checks, these results show the advantages of DIDA's generated samples from both rule-based and statistical perspectives.

Table: Statistical Diagnostics of Generated Samples (%, Mean ± Sample Standard Deviation)

DatasetDiagnosticOriginalSMOTECTGANSMOTE-CDNNTabDDPMMDLMTANDEMDIDA
AlzheimersNumeric range0.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.0000
Joint support0.0000 ± 0.00002.230 ± 0.84132.1308 ± 0.90501.8138 ± 0.31791.8894 ± 1.2557<u>1.5713 ± 0.7634</u>3.8343 ± 0.46650.0019 ± 0.0004
KidneyNumeric range0.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.0000
Joint support0.0000 ± 0.0000<u>0.1436 ± 0.0016</u>0.5373 ± 0.02430.6112 ± 0.00310.8876 ± 0.00310.1623 ± 0.00221.0975 ± 0.03630.0047 ± 0.0005
HeartNumeric range0.3325 ± 0.15350.2362 ± 0.16102.7007 ± 1.1521<u>0.1884 ± 0.1337</u>0.0000 ± 0.00000.0000 ± 0.00000.7746 ± 0.34980.0000 ± 0.0000
Joint support0.0000 ± 0.0000<u>1.2432 ± 0.3908</u>2.1827 ± 0.81381.2495 ± 0.45811.6193 ± 0.52092.0902 ± 1.78032.5272 ± 1.67620.0533 ± 0.0057
HealthcareNumeric range0.0255 ± 0.01440.0082 ± 0.00650.0369 ± 0.0671<u>0.0046 ± 0.0052</u>0.0000 ± 0.00000.0000 ± 0.00001.6400 ± 0.62210.0000 ± 0.0000
Joint support0.0000 ± 0.00001.5761 ± 0.0731.6190 ± 0.03881.0040 ± 0.06181.2119 ± 0.2070<u>0.8319 ± 0.0062</u>1.1918 ± 0.24800.0010 ± 0.0012
TeenNumeric range0.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.0000
Joint support0.0000 ± 0.00005.0325 ± 0.91656.6819 ± 1.00415.9841 ± 1.47386.1307 ± 1.9614<u>4.3365 ± 0.8287</u>7.041 ± 1.79840.0000 ± 0.0000
UrbanNumeric range0.0000 ± 0.00000.0000 ± 0.00000.0025 ± 0.00460.0000 ± 0.00000.0000 ± 0.00000.0000 ± 0.0000<u>0.0008 ± 0.0022</u>0.0000 ± 0.0000
Joint support0.0000 ± 0.000010.9432 ± 1.747011.2495 ± 1.488210.9217 ± 2.003910.7888 ± 2.9394<u>9.4224 ± 1.7022</u>13.0311 ± 2.44570.7451 ± 0.3351
Reviewer_B7a7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper aims to augment data while satisfying compliance, auditability and differentiation. It proposes DIDA, an effective method which computes data importance automatically and carries out customized augmentation according to the constraints among features. Theoretical analysis shows that data with higher importance should be less perturbated, backing up DIDA. The main contributions of this paper mainly lie in the combination of theoretical findings and empirical results of the data augmentation framework.

Critical strengths
  1. This paper proposes a novel framework integrating theory and experiments. Starting from the requirements of data augmentation, this paper studies the relationship between the sample importance and the corresponding perturbation intensity in theory. The proposed DIDA pipeline evaluates the importance, investigates the constraints and conducts differential augmentation. Extensive experiments across diverse datasets highlight the effectiveness of DIDA.
  2. This paper digs into the mechanism and does not blindly apply LLM to solve the problem. As a prevailing base, LLM suffers from slow reasoning process and might not be suitable for data augmentation which needs more outputs. This paper investigates from the perspective of data and does not incorporate LLM into the augmentation process, making the pipeline relatively light-weighted and capable for different domains with limited computing resources.
Critical weaknesses
  1. More experiments can be added to highlight DIDA's advantage. DIDA does not use LLM, probably faster than other LLM-based methods. Time complexity analysis can be carried out to find the upper bound of DIDA's execution time. Augmentation time comparison can also be conducted to show DIDA's performance more comprehensively. Moreover, the datasets are of different sizes and a sensitivity analysis regarding the dataset size can be carried out to compare the performance of different methods.
  2. Some figures in this paper can be further improved. The text in the middle of Figure 1 is too small to recognize. The colors used by the legends in Figure 2 may be hard to tell. The (c) and (d) subfigures in Figure 4 have collision in the tick label. The legend SMOTE and CDNN seem identical in Figure 5.
Questions for authors
  1. Corresponding to the Weakness section, the proposed method does not use deep models and may exhibit short time consumption. What is the upper bound of DIDA's time cost? Given the performance improvement, can we say that DIDA is both effective and efficient?
  2. Corresponding to the Weakness section, what is the result for DIDA's performance when dataset size varies? Does the performance gap gets bigger when the dataset gets smaller?
3 comments

W1 / Q1 Time complexity and runtime efficiency

Following your suggestion, we have added an augmentation time comparison. We discuss both runtime and predictive performance to give a more complete picture of DIDA's performance.

(1) We compared the generation time of each method using 100% of the training data, as shown in the table below. Although DIDA requires data valuation and rule construction, its generation time is still lower than that of most comparison methods.

(2) Compared with SMOTE, DIDA takes more time. We add valuation and rule construction in the hope of allocating augmentation according to the task value of samples and features. Considering the existing predictive results and the supplementary runtime measurements together, we believe that DIDA achieves a good balance between augmentation effectiveness and runtime efficiency in the current settings.

Table: Generation time using 100% of the training data (seconds)

DatasetDIDASMOTECTGANSMOTE-CDNNTabDDPMMDLMTANDEM
Alzheimers30.068 ± 9.7307.162 ± 0.00370.025 ± 1.83129.870 ± 1.40837.548 ± 0.33648.499 ± 1.58836.902 ± 0.291
Kidney27.687 ± 21.0198.221 ± 0.00878.085 ± 3.00528.884 ± 1.89536.779 ± 0.31945.951 ± 1.96435.857 ± 0.173
Heart19.960 ± 0.3597.560 ± 0.00537.004 ± 0.40824.012 ± 1.02034.273 ± 0.08541.487 ± 0.38534.138 ± 0.194
Healthcare48.756 ± 4.20115.469 ± 0.035307.270 ± 20.01385.156 ± 14.75275.338 ± 1.236137.546 ± 2.07968.503 ± 1.146
Teen20.451 ± 1.4703.069 ± 0.00341.502 ± 0.71223.794 ± 0.93035.106 ± 0.22541.143 ± 2.18734.436 ± 0.120
Urban29.887 ± 2.9498.275 ± 0.00496.723 ± 2.35146.929 ± 5.06243.239 ± 0.47461.038 ± 0.51041.334 ± 0.180

W1 / Q2 Training set size and performance with limited data

To answer more clearly whether DIDA is effective with limited data, we have added experiments at different training set sizes and explain them in terms of the experimental setup and the analysis of the results.

(1) Following your suggestion, we have added comparisons at different training set sizes within each dataset, using 25%, 50%, 75%, and 100% of the original training set to compare the methods' ROC-AUC, AP, and F1 scores. Due to the limited space in this response, we present only ROC-AUC here, and the results for the other two metrics will be added to the paper.

(2) As the table shows, DIDA achieves the highest mean ROC-AUC for most combinations of training set size and dataset, and also maintains its lead at the smallest training fraction on all six datasets. We also found that, on some datasets, DIDA's advantage over the strongest comparison method grows further as the number of training samples decreases. This shows that, on the tested datasets, our method can still achieve good predictive performance with less training data.

Table: ROC-AUC at different training fractions (mean ± sample standard deviation)

DatasetTrain (%)OriginalSMOTECTGANSMOTE-CDNNTabDDPMMDLM-localTANDEMDIDA
Alzheimers25%<u>0.8224 ± 0.0075</u>0.8183 ± 0.00880.7870 ± 0.01230.8171 ± 0.00380.7901 ± 0.00970.8077 ± 0.01330.8107 ± 0.00880.8429 ± 0.0094
50%0.8507 ± 0.00960.8517 ± 0.01210.8314 ± 0.0075<u>0.8535 ± 0.0091</u>0.8301 ± 0.01080.8377 ± 0.00840.8378 ± 0.00590.8654 ± 0.0080
75%<u>0.8630 ± 0.0078</u>0.8617 ± 0.00850.8497 ± 0.01100.8606 ± 0.00990.8447 ± 0.00820.8514 ± 0.00800.8529 ± 0.00690.8708 ± 0.0069
100%<u>0.8672 ± 0.0076</u>0.8654 ± 0.01040.8538 ± 0.00680.8656 ± 0.00750.8508 ± 0.01010.8580 ± 0.00840.8577 ± 0.00910.8745 ± 0.0062
Kidney25%0.6332 ± 0.03130.6398 ± 0.02950.6150 ± 0.03850.6434 ± 0.03150.5933 ± 0.03410.6434 ± 0.0247<u>0.6457 ± 0.0348</u>0.6639 ± 0.0418
50%0.6765 ± 0.03110.6776 ± 0.02710.6495 ± 0.0347<u>0.6837 ± 0.0280</u>0.6244 ± 0.03410.6720 ± 0.02850.6741 ± 0.02050.6945 ± 0.0246
75%0.6939 ± 0.0252<u>0.6998 ± 0.0299</u>0.6609 ± 0.03440.6974 ± 0.03000.6471 ± 0.03420.6930 ± 0.02910.6878 ± 0.04020.7099 ± 0.0281
100%0.7050 ± 0.01670.7100 ± 0.02570.6770 ± 0.02420.7149 ± 0.01900.6589 ± 0.03050.7129 ± 0.03380.6939 ± 0.0172<u>0.7136 ± 0.0187</u>
Heart25%0.8203 ± 0.01530.8127 ± 0.00990.8131 ± 0.02100.8157 ± 0.01470.8261 ± 0.01830.8201 ± 0.0141<u>0.8270 ± 0.0196</u>0.8306 ± 0.0170
50%0.8392 ± 0.01710.8262 ± 0.02260.8277 ± 0.01870.8324 ± 0.0227<u>0.8398 ± 0.0161</u>0.8368 ± 0.02190.8328 ± 0.02010.8489 ± 0.0236
75%0.8453 ± 0.02020.8397 ± 0.02340.8439 ± 0.02670.8393 ± 0.0206<u>0.8472 ± 0.0229</u>0.8414 ± 0.01250.8405 ± 0.02060.8574 ± 0.0247
100%0.8419 ± 0.01290.8416 ± 0.0200<u>0.8621 ± 0.0257</u>0.8442 ± 0.01850.8605 ± 0.03030.8521 ± 0.03430.8518 ± 0.02050.8725 ± 0.0294
Healthcare25%<u>0.7410 ± 0.0101</u>0.7403 ± 0.00700.7402 ± 0.00780.7385 ± 0.00730.7397 ± 0.00480.7285 ± 0.00680.7311 ± 0.00710.7469 ± 0.0052
50%<u>0.7514 ± 0.0065</u>0.7478 ± 0.00680.7489 ± 0.00670.7460 ± 0.00750.7487 ± 0.00640.7387 ± 0.00680.7393 ± 0.00800.7548 ± 0.0059
75%<u>0.7539 ± 0.0061</u>0.7519 ± 0.00560.7516 ± 0.00570.7500 ± 0.00530.7523 ± 0.00510.7441 ± 0.00510.7409 ± 0.00720.7573 ± 0.0045
100%<u>0.7554 ± 0.0058</u>0.7526 ± 0.00490.7529 ± 0.00630.7514 ± 0.00400.7537 ± 0.00440.7471 ± 0.00520.7422 ± 0.00690.7585 ± 0.0050
Teen25%0.8237 ± 0.04820.8369 ± 0.04040.8267 ± 0.04090.8453 ± 0.04910.7036 ± 0.06030.8474 ± 0.0336<u>0.8521 ± 0.0563</u>0.8554 ± 0.0360
50%0.9025 ± 0.0469<u>0.9213 ± 0.0467</u>0.8876 ± 0.03730.9403 ± 0.03920.7223 ± 0.04770.8918 ± 0.04750.9054 ± 0.02860.9151 ± 0.0400
75%0.9143 ± 0.03770.9318 ± 0.03850.9350 ± 0.0212<u>0.9373 ± 0.0249</u>0.7867 ± 0.06410.9251 ± 0.03000.9286 ± 0.03260.9511 ± 0.0268
100%0.9398 ± 0.02030.9514 ± 0.02670.9285 ± 0.0281<u>0.9589 ± 0.0180</u>0.8071 ± 0.04440.9413 ± 0.02170.9458 ± 0.03610.9610 ± 0.0303
Urban25%<u>0.5926 ± 0.0065</u>0.5832 ± 0.01050.5856 ± 0.00820.5812 ± 0.01050.5790 ± 0.00780.5794 ± 0.00940.5777 ± 0.00920.5964 ± 0.0045
50%<u>0.6069 ± 0.0037</u>0.6030 ± 0.00530.5988 ± 0.00590.6011 ± 0.00750.6008 ± 0.00840.6021 ± 0.00650.5912 ± 0.00450.6113 ± 0.0054
75%<u>0.6201 ± 0.0031</u>0.6089 ± 0.00510.6091 ± 0.00400.6081 ± 0.00410.6072 ± 0.00250.6094 ± 0.00520.6029 ± 0.00560.6250 ± 0.0034
100%<u>0.6240 ± 0.0036</u>0.6131 ± 0.00550.6112 ± 0.00490.6151 ± 0.00660.6144 ± 0.00660.6167 ± 0.00400.6110 ± 0.00700.6277 ± 0.0041

W2 Figure readability and visual distinction between methods

We will address the figure details you identified through the following four changes, making the key text and the differences between methods easier to see.

(1) To address the small text in the middle of Figure 1, we will enlarge the key text, simplify secondary annotations that affect readability, and adjust the information density to make the main workflow and key content easier to identify.

(2) In Figure 2, we will use more distinct legend colors together with shapes or text labels, so that readers can distinguish the content with less reliance on color alone.

(3) To address the overlapping tick labels in Figures 4(c) and 4(d), we will adjust their spacing and arrangement so that the axis information is clearly readable.

(4) In Figure 5, we will focus on adjusting the fills and textures for SMOTE and CDNN and keep the legends consistent with their corresponding boxplots, so that the two methods do not look the same.

In the revised PDF, blue text and margin labels identify revisions made in response to your comments (Reviewer B7a7). Please refer to the revised PDF and supplementary material for details.

We welcome further discussion and would be happy to clarify any remaining questions.

Reviewer_3EkMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose DIDA, a data augmentation framework that uses feature importance and sample contribution to determine perturbation intensity. The central idea is interesting and timely, given the growing interest in data augmentation.

Critical strengths
  1. The framework is clean and intuitive, and adapting augmentation intensity to feature importance and sample contribution is a natural and well-motivated approach.

  2. The experimental results are promising and interesting, showing consistent improvements over multiple datasets.

Critical weaknesses
  1. The presentation is not clear enough. For example, the first paragraph of the introduction mainly focuses on hard constraints and class imbalance, but their connection to the auditability and differentiation requirements introduced in the Challenges is not clear enough. The motivation for Challenge 1 also needs explanation. How do sample and feature perturbation bounds address the aforementioned problems?

  2. Please see my questions below.

Questions for authors
  1. Boundary constraints in Theorem 1: Definition 4 introduces constraints on feature values, but these constraints do not appear in the KKT formulation. Is it implicitly assumed that the contraints are not active? In that case how reasonable is that assumption?

  2. Assumption in Theorem 1: What is the intuition of δij≪∣xij∣\delta_{ij}\ll |x_{ij}| in Theorem 1? It seems that shifting the data by a large constant c can make this condition easier to satisfy, even though the perturbation remains unchanged after mapping back to the original coordinates.

  3. Connection between theory and method: Why does Definition 5 use the function in Equation (4), instead of the expression in Theorem 1? Also the algorithm does not impose the cosine-similarity constraint used in the analysis. Can you clarify to which extent the theorem supports the practical method?

  4. Intuition of the building the constraint table: Can the authors clarify why they choose to build such table with discretized numerical features, instead of more flexible approaches such as decision trees?

6 comments

Thank you for carefully reviewing our paper and recognizing the intuitive nature of the DIDA framework, the motivation for value-guided augmentation, and the promise of the experiments across multiple datasets. Regarding your questions about the connections in the introduction, theoretical assumptions, the relationship between theory and the practical method, and the design of the conditional tables, we will further explain our design considerations and improve the unclear parts of the paper.

W1 Background, differentiation, and auditability in the introduction

To address the unclear connection you identified between the application background and the subsequent discussion, we plan to revise the introduction in terms of the analysis of requirements and the connections between the preceding and subsequent discussion, so that readers can more easily understand why we designed DIDA this way.

(1) With limited data, augmentation needs to make full use of the available training information; with class imbalance, its allocation needs careful consideration; and in settings with strong constraints, generated changes need to be controlled and their basis explained. We will analyze these requirements more fully in the introduction and connect them to the design of DIDA when presenting the problems in the Challenges section, bringing the background, method design, and technical challenges into closer alignment.

(2) In addition, we will provide appropriate context earlier in the text for the differentiated augmentation and auditability discussed later, explaining why these designs are needed and the respective roles of perturbation intensity, value boundaries, and conditional tables. This will allow readers to understand how these designs relate to the preceding requirements when they reach the later discussion, avoiding the overly abrupt introduction of concepts and mechanisms.

Q1 Value boundaries in Definition 4 and the KKT derivation

We explain the value constraints in Definition 4 from three perspectives: the purpose of the derivation, the effect of the boundaries, and how they are handled during generation.

(1) The current derivation corresponds to the case where the value boundaries are inactive. We derive the theorem to analyze how sample value and feature value affect each cell's perturbation magnitude under a given consistency requirement, so it is necessary to separate the effect of value from the additional restrictions imposed by boundaries.

(2) For example, if a cell with low combined value is close to the upper bound, its available upward range is restricted even if the value-based allocation principle permits a large perturbation. If the resulting perturbation is small, it is then difficult to distinguish whether this comes from value-based allocation or the boundary restriction. Temporarily excluding the upper and lower bounds makes the relationship between perturbation and value easier to analyze; the corresponding closed-form analysis applies when the boundaries are inactive.

(3) In actual augmentation, we first determine the perturbation magnitude based on value and then apply boundary handling to constrain candidates that fall outside the bounds. This allows us to allocate perturbation space according to value while keeping the final values within the corresponding ranges.

Q2 The small-perturbation assumption and translation of the data

Regarding the translation issue, we would like to explain two points: the meaning of the small-perturbation assumption and which quantities change after translation.

(1) In Theorem 1, δij≪∣xij∣\delta_{ij}\ll|x_{ij}| means that the perturbation is sufficiently small relative to the original value in the current feature representation. We introduce this assumption to make a second-order approximation of the weighted cosine around the original sample and analyze how sample value and feature value affect perturbation allocation.

(2) Translating a given original sample and its augmented counterpart by the same amount does not change their difference. When the translation increases the absolute values, the ratio of the perturbation to the absolute value after translation becomes smaller, but the actual change does not, so we cannot conclude from this that the sample has been modified less. If we further recompute the theoretical solution using the translated data, the weighted cosine and Ai=∑jbj′xij2A_i=\sum_jb'_jx_{ij}^2 generally change, and so may the theoretical budget and the resulting perturbation magnitude. Therefore, this assumption needs to be understood in the context of the chosen feature representation.

Q3 The theorem, the practical perturbation function, and the cosine constraint

We explain the relationship between the theorem and the practical method through three points: the design of Equation (4), the use of the cosine constraint, and what the supplementary experiments show.

(1) The proof of Theorem 1 is intended to analyze how data value affects perturbation allocation under a given consistency requirement and guide the design of the practical method. Equation (4) puts this relationship into practice: it retains the property that perturbation intensity does not increase as value increases, bounds the intensity range, avoids excessively large perturbations when feature importance approaches zero, and allows the parameters to be adjusted to the characteristics of different datasets.

(2) The cosine-similarity constraint is primarily used in mathematical modeling to help us analyze the relationship between the consistency requirement and perturbation allocation. Whether to include this constraint in actual augmentation and how to set its threshold still need to be considered in relation to the specific task, so we have not imposed a uniform setting in the general pipeline, keeping the implementation simple.

(3) We added a direct comparison of inverse, uniform, and forward perturbation, with the results shown in the table below. Inverse perturbation achieves the highest mean on all three metrics across the six datasets, indicating that, under the current experimental settings, prioritizing the protection of high-value information produces better predictive performance than uniform or forward allocation. These results support the allocation direction proposed by the theory and also show that it can effectively guide the design of the practical perturbation function.

Table: Predictive performance under different perturbation directions

DatasetDirectionROC-AUC (%) ↑AP (%) ↑F1 (%) ↑
AlzheimersInverse (DIDA)87.45 ± 0.6280.98 ± 1.2376.84 ± 1.44
Uniform86.27 ± 0.8978.87 ± 1.6073.69 ± 2.01
Forward83.42 ± 0.6774.83 ± 1.1569.16 ± 1.98
Chronic KidneyInverse (DIDA)71.36 ± 1.8795.56 ± 0.4695.28 ± 0.28
Uniform70.08 ± 1.7195.15 ± 0.3694.89 ± 0.17
Forward68.71 ± 2.4495.02 ± 0.5294.80 ± 0.32
HeartInverse (DIDA)87.25 ± 2.9487.25 ± 2.9883.93 ± 2.38
Uniform85.54 ± 2.4885.76 ± 2.5982.76 ± 2.06
Forward86.33 ± 2.7786.10 ± 2.8783.02 ± 2.21
HealthcareInverse (DIDA)75.85 ± 0.5070.34 ± 0.3864.43 ± 1.14
Uniform74.86 ± 0.5369.38 ± 0.4563.90 ± 1.40
Forward74.87 ± 0.5369.39 ± 0.4163.83 ± 1.29
TeenInverse (DIDA)96.10 ± 3.0377.64 ± 9.4161.64 ± 9.47
Uniform94.49 ± 2.6775.29 ± 8.0355.22 ± 10.86
Forward92.84 ± 3.6263.99 ± 12.0633.22 ± 9.97
Urban StreetInverse (DIDA)62.77 ± 0.4155.59 ± 0.3347.49 ± 3.22
Uniform61.24 ± 0.4054.93 ± 0.2944.99 ± 3.99
Forward60.91 ± 0.5554.69 ± 0.4346.13 ± 3.36

Q4 Design choices for discretization and conditional probability tables

We also considered decision trees during the design process and ultimately chose discretization and conditional probability tables mainly for the following two reasons.

(1) Conditional partitions after discretization are easier to control and also make it easier to check the basis for categorical replacement. We can combine numerical intervals with categorical conditions and then estimate the probability of each candidate category from the training samples that satisfy those conditions. In this way, for an exact match, we can directly check under which conditions and on the basis of which statistical results the replacement was made.

(2) We also considered the need to incorporate expert knowledge later. If experts provide explicit business thresholds, the bins can be adjusted accordingly so that the conditional partitions better reflect known business boundaries, and category probabilities can be estimated from the training data under these conditions. Although decision trees can partition the data adaptively, additional constraints would be needed to make the learned partitions follow the same expert-defined boundaries.

Of course, binning also loses detail within intervals, and some combinations of conditions may have no corresponding training observations, so we include a fallback procedure to handle these cases. In the revision, we will further explain this choice and the effects of binning granularity and sparse conditions.

In the revised PDF, red text and margin labels identify revisions made in response to your comments (Reviewer 3EkM). Please refer to the revised PDF and supplementary material for details.

Thank you for the responses and additional experiments. I will maintain my original score for now and discuss it further during the AC reviewer discussion.

Comments on the paper

1 comment

Thank you to all four reviewers for the time and care you put into the paper. 3EkM’s translation example helped us clarify the small-perturbation assumption. B7a7’s question about training size prompted us to compare methods at four training fractions. jjVv suggested testing three perturbation directions directly, and PQ1V asked whether generated records still satisfy domain rules when sampling conditions are relaxed. We have not yet heard further from B7a7 or jjVv. For the reviewers and the Area Chair, we summarize the revisions and experiments below.

1. Motivation and theory (3EkM W1, Q1–Q3; jjVv W1–W2, Q1–Q2)

  • Motivation (Sec. 1, Fig. 1). The revised introduction explains why class counts alone do not determine how much to change each value. DIDA uses sample and feature scores to do so.
  • Derivation (Sec. 3.2, Thm. 1; App. D, Eqs. 31–32). For small perturbations, positive weights, and inactive bounds, magnitude is proportional to the reciprocal of feature weight within a sample (fixed weight vector). The weight is outside the square root. As 3EkM noted, a translation can change the relative perturbation without changing the increment.
  • Implementation (Sec. 4, Fig. 2; Def. 5, Eq. 4). Both scores guide numerical and categorical changes. Eq. (4) bounds the strength and keeps it from rising with value. The cosine constraint is used in the derivation.
  • Direct test (Sec. 5.3, Fig. 7; App. F.7.3, Table 15). On Alzheimers, mean F1 was 76.84% for inverse allocation, versus 73.69% for uniform and 69.16% for proportional allocation.

2. Limited training data (B7a7 W1, Q2; PQ1V W1)

  • Range correction (Sec. 4.1, Fig. 3). Bounds based on the interquartile range can exclude rare predictive values in small samples. Sample contribution guides range correction; we tested identifiable domain rules separately.
  • Training fractions (Sec. 5.2, Fig. 5; App. F.7.2, Tables 12–14). We used 25%, 50%, 75%, and 100% of each training set. DIDA achieved the highest mean ROC-AUC in 22 of 24 comparisons, including all six datasets at 25%.
  • Size of the advantage (App. F.7.2, Table 12). B7a7 asked whether the gap widens as training data become scarcer. On Alzheimers, DIDA’s mean ROC-AUC lead over the best comparator rose from 0.73 to 2.05 percentage points as training data fell from 100% to 25%.

3. Direction and scores (3EkM Q3; jjVv W4, Q4)

  • Direction (Sec. 5.3, Fig. 7; App. F.7.3, Table 15). We varied only how numerical perturbation strength depends on value. Inverse allocation gave the highest mean ROC-AUC, AP, and F1 on each of the six datasets.
  • Score assignment (Sec. 5.3, Table 1). Shuffling only feature scores lowered macro-average F1 by 2.98 percentage points; shuffling only sample scores reduced it by 1.16 points. Original scores ranked first on all three metrics.
  • Robustness (Sec. 5.3, Table 1; App. F.7.3). With score noise of 0.10, 0.25, or 0.50 SD, or with random forest valuation, all three macro-averages exceeded the no-augmentation baseline.

4. Constraints and fallback (3EkM Q4; jjVv W3, Q3; PQ1V W1–W2)

  • Conditional tables (Sec. 4.2; App. A.2.4; initial response to 3EkM Q4). Explicit bins make exact-match counts traceable and allow expert cutoffs. If relaxing conditions yields no match, DIDA samples from the marginal distribution or, if unavailable, uniformly from the feature domain.
  • Domain rules (rule tables in our initial responses to jjVv and PQ1V). We checked blood pressure ordering in Alzheimers, Kidney, and Healthcare, and BMI consistency in Healthcare. Reported violations were 0% in all four checks on final outputs.
  • Joint support (App. F.7.5, Table 17). For each sample–feature-group check, we asked whether its tuple occurred in the filtered training reference. DIDA had the lowest absence rate and run-to-run SD on all six datasets among augmentation methods. We plan further checks after range correction.

5. Runtime and figures (B7a7 W1, Q1, W2)

  • Runtime (App. F.7.4, Table 16). At full training size, DIDA generated data faster than CTGAN, TabDDPM, MDLM, and TANDEM on all six datasets, counting valuation and rule construction. It was faster than SMOTE-CDNN on five; SMOTE was fastest.
  • Figures (initial response to B7a7 W2; original Figs. 1, 2, 4(c)–(d), 5). We proposed larger labels, clearer legends and textures, and wider tick spacing. Revised Fig. 5 shows training fractions.

In the revised PDF, the text and margin labels are color-coded by reviewer: red for 3EkM, blue for B7a7, green for jjVv, and purple for PQ1V. Thank you again for the time you spent on the derivations and experiments. Your questions helped us explain why we made these choices and decide which comparisons would be most informative. We would welcome a fuller discussion, whether you have a question about a particular result or a different reading of the evidence.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.