Building a stable classifier with the inflated argmax

We propose a new framework for algorithmic stability in the context of multiclass classification. In practice, classification algorithms often operate by first assigning a continuous score (for instance, an estimated probability) to each possible label, then taking the maximizer -- i.e., selecting the class that has the highest score. A drawback of this type of approach is that it is inherently unstable, meaning that it is very sensitive to slight perturbations of the training data, since taking the maximizer is discontinuous. Motivated by this challenge, we propose a pipeline for constructing stable classifiers from data, using bagging (i.e., resampling and averaging) to produce stable continuous scores, and then using a stable relaxation of argmax, which we call the "inflated argmax," to convert these scores to a set of candidate labels. The resulting stability guarantee places no distributional assumptions on the data, does not depend on the number of classes or dimensionality of the covariates, and holds for any base classifier. Using a common benchmark data set, we demonstrate that the inflated argmax provides necessary protection against unstable classifiers, without loss of accuracy.

Paper

References (35)

Scroll for more · 23 remaining

Similar papers

Peer review

Reviewer LzW37/10 · confidence 3/52024-07-08

Summary

This paper studies and proposes a new theoretical framework to study algorithmic stability for multiclass classification with a focus on a stable class selection given predicted class probability. Based on the proposed theory, the proposed selection criterion based on bagging called "inflated argmax" was proposed. Theoretical analysis suggests that the proposed inflated argmax is stable, while also give a relatively tight decision for the candidate prediction set. Simple experiments using Fashion-MNIST were also provided to show the stability of inflated argmax as well as how tight the candidate set can be.

Strengths

1. The paper provides strong theoretical framework for future study of stable selection criteria for the multiclass classification problem given predicted class probability. 2. The proposed notion of $\epsilon$- compatible is simple to understand and a reasonable extension of classical algorithmic stability noted in Definition 2. I feel this notion is neat and also very useful for further analysis in Proposition 4 and beyond. 3. The simple rule that satisfies the stability but could give a looser quality than inflated argmax was also provided to give a better understanding of the problem. The study was quite detailed in Appendix D., which I find quite amazing. 4. Despite the complicated notion of inflated argmax, the efficient computation of selection rule once bagging is done is nicely provided in Proposition 10.

Weaknesses

1. Experiments were quite weak in my opinion to highlight the effectiveness of inflated argmax. For example, more reasonable baseline that has larger than 1 $\beta_{\mathrm{size}}$ could have been provided to show that inflated argmax is quite favorable if it allows to predict more than one class. Curreent baselines are restrictive in the sense that only one class can be chosen (argmax). Several improvements could also be done (please see the questions section for possible improvements). 2. The proposed inflated argmax requires bagging, which can be quite computationally expensive, or prohibitive in some scenarios under limited resources or large dataset/model.

Questions

1. Can we also see the performance of inflated argmax *without bagging* in the experiment? In my understanding, this should also be possible with the proposed efficient algorithm. This can be useful to see how important the bagging method is needed for the proposed algorithm. 2. Visualizing the number of bags needed to perform reasonably in the experiment can also be useful. Because it can be expensive if we use more sophisticated algorithm or using larger dataset. 3. In deep learning, overconfident problem of probability can often be an issue when one wants to use probability score. How does this problem affect the proposed inflated argmax in practice? 4. Can we know that it is likely that the second class to be chosen as a candidate class is likely to be the second-largest probability class? 5. Is it reasonable to say that precision is similar given different average precision and standard error provided (Lines 293--294)? 6. I was wondering if there are other possible baselines in related work that could be used in the experiments? 7. Suggestions on simple baselines 7.1 Top-2 class prediction, size of beta will be 2. 7.2 Predicting classes until sum of probability becomes a certain constant, e.g., if we set threshold = 0.8 and prediction probability vector is [0.1, 0.1, 0.1, 0.2, 0.6], then we will pick the last two classes (0.2 and 0.6). This simple idea looks intuitive and gives flexible set and this is the most simple baseline I had in mind that could be effective when I first read this paper. How do you think about this baseline? I think the core contribution of this paper is the theory of stable classification rule. Therefore, although the experiment looks quite weak, my current rating is on the accept side (weak accept). I know it can be too much to ask to improve experiments a lot and I do not expect authors to fix everything according to my suggestions. But I would appreciate if the authors can reply to make sure that my understanding is correct.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Discussion of limitations of the proposed method and potential future work were appropriate.

Reviewer m1xX7/10 · confidence 4/52024-07-09

Summary

An algorithm is considered stable if small perturbations in the training data do not result in significant changes to its output. Stability has been previously explored in the context of regression. This paper extends this concept to the multiclass framework, where previous work has focused on stability in terms of output probability rather than label selection processes. Given that argmax is not a continuous function, it can lead to instability in the classifier. To address this issue, the authors first define stability for the multiclass domain and then demonstrate that a stable classifier alone is insufficient; a corresponding label selection strategy is also necessary. The authors propose a framework combining bagging and inflated argmax to derive such a combination, which they validate experimentally, showing that their approach achieves higher stability while maintaining precision at a minimal cost.

Strengths

1. The authors introduce a novel definition of algorithmic stability in classification, shifting the focus from output probability to predicted labels. They demonstrate that, according to this definition, a stable classifier is insufficient for achieving overall stability. Furthermore, they establish a connection between this notion of stability and classical algorithmic stability concepts. Building on these insights, the authors propose an approach that leverages bagging and inflated argmax to achieve said stability. 2. It is essential to emphasize that the proposed framework operates without assuming any specific distribution on the data. Additionally, it does not rely on the number of classes or dimensionality of the covariates, making it versatile and applicable to various algorithms. 3.The experimental findings demonstrate the stability of the proposed framework. Moreover, they reveal the negative impact of argmax on stability. Notably, the results show that the proposed approach incurs only a minor penalty to precision, underscoring its effectiveness.

Weaknesses

1.The authors point out my primary concern in their discussion section, highlighting the practicality of the proposed framework. Notably, bagging is an computationally expensive procedure, and while the authors suggest that parallelization can mitigate this issue, I agree that this does not necessarily translate to a reduction in overall cost. This becomes particularly challenging when dealing with large datasets commonly encountered in modern classifiers (e.g., image classification), where the sheer scale of computation required can be prohibitively expensive even with multiple GPUs. 2. The experimental section relies heavily on a single dataset and classifier combination. To fully demonstrate the benefits of the proposed framework, I believe it would be necessary to conduct a more comprehensive set of experiments. Specifically, I suggest that the authors provide additional evidence through experimentation to illustrate how the slight tradeoff in precision is justified by the gained stability in practice.

Questions

1. From the definition of stability given in 4, it is evident that the authors trained the classifiers multiple times while removing training samples. How does the overall average of average precision look like in such cases? Such a metric might show the advantage the proposed framework. 2. My other main question is related to practicality. I would like to know the author's thought more regarding that limitation.

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

N/A

Reviewer sBU36/10 · confidence 3/52024-07-10

Summary

This paper focuses on the stability of learning multiclass classfiers. It proposes a notion of selection stability for the learning algorithm as well as a modified argmax that returns a set of labels. Evaluations have been conducted using FashionMNIST and simple models.

Strengths

- This paper focuses on the problem of classifier stability, which has high practical interest. - This paper proposes a relaxed form of argmax, which contributes concretely to the problem. - Propositions and theoretical relations to existing results are soundly presented.

Weaknesses

- The presentation of the proposed contribution can be improved. - The core proposal of inflated argmax is introduced in section 3.2, which is too late. - Bagging is not part of proposal, and shows low relation to the contribution, which can be demonstrated using fewer space. - Additional properties of the inflated argmax is also not directly related to the problem. - "learning stability by leave-one-out data" and "prediction stability when using argmax for similar probabilities" seem to be mixed and not clearly addressed. - Some sentences such as `workflow is fundamentally incompatible with the goal of algorithmic stability` is too extreme, for example, simple streament using proper regularization terms can somehow mitigate the problem. - The demonstration of empirical evaluation does not match the last sentence in abstract. - The combination of $\epsilon$-inflated argmax on the base learning algorithm should be compared. - It is also nice to show results on at least two datasets.

Questions

- How tail stability and selection stability relates to $\epsilon$-compatibility differently? To confirm, do bagging classfiers only satisfy tail stablity? Are there other algorithm designs satisfy these stabilities? - How to better interpreter the empirical evaluation results. With similar accuracy shown in Table 1, the proposed method should show better stability in Figure 2?

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

2

Limitations

Practical limitations are presented in section 5.

Reviewer FNaQ7/10 · confidence 3/52024-07-10

Summary

This submission studies the problem of making set-valued predictions, in which the classifier should strive for an optimal balance between the correctness (the true class is among the candidates) and the precision (the candidates are not too many) of its prediction. Yet, this submission mentions one method for this purpose [1], which predicts, for each query instance, a Bayes-optimal set valued prediction optimizing the F-measure. It seems to miss a discussion on (several) related methods, which are constructed based on both the probability theory and imprecise probability, e.g., the ones discussed in [3, 4, 6] and references therein. On the abstract level, the inflated argmax presented in Section 3.2 might be seen as a way to distort/imprecise the given singleton conditional probability distribution. It might enlarge the set of existing approaches targeting similar purposes, such as the ones discussed in [2, 5] and references therein, meaningfully. The proposed algorithm is compared with a variant of LeNet-5, implemented in PyTorch [PGML19] tutorials as GarmentClassifier(), and its ensemble version on the Fashion-MNIST data set. Empirical evidence suggests that the proposed algorithm provides better $\beta_{\text{prec}}$ than the single classifier, but worse than the ensemble version. The average set size produced by the proposed algorithm seems to be reasonably small. Results on the selection instability $\delta_j$ defined in (4), i.e., an approximate of the loss version of the stability $\delta$ defined in (1), which essentially measures the robustness of the set-valued predictions under tiny changes in the training data set, seems to be in favor of the proposed algorithm. [1] Del Coz, J. J., Díez, J., & Bahamonde, A. (2009). Learning Nondeterministic Classifiers. Journal of Machine Learning Research, 10(10). [2] Montes, I., Miranda, E., & Destercke, S. (2020). Unifying neighbourhood and distortion models: part I–new results on old models. International Journal of General Systems, 49(6), 602-635. [3] Mortier, T., Wydmuch, M., Dembczyński, K., Hüllermeier, E., & Waegeman, W. (2021). Efficient set-valued prediction in multi-class classification. Data Mining and Knowledge Discovery, 35(4), 1435-1469. [4] Nguyen, V. L., Destercke, S., Masson, M. H., & Hüllermeier, E. (2018, July). Reliable multi-class classification based on pairwise epistemic and aleatoric uncertainty. In 27th International Joint Conference on Artificial Intelligence (IJCAI 2018) (pp. 5089-5095). [5] Nguyen, V. L., Zhang, H., & Destercke, S. (2023, September). Learning Sets of Probabilities Through Ensemble Methods. In European Conference on Symbolic and Quantitative Approaches with Uncertainty (pp. 270-283). Cham: Springer Nature Switzerland. [6] Troffaes, M. C. (2007). Decision making under uncertainty using imprecise probabilities. International journal of approximate reasoning, 45(1), 17-29.

Strengths

S1: I think taking into account the instability, due to tiny changes in the training data set, when making set-valued prediction is interesting. Yet, assessing this aspect of classifiers seems to be costly, i.e., requiring a sufficiently large number of leave-one-out models. S2: Empirical evidence seems to be promising. For example, $\beta_{\text{prec}}$ seems to suggest that the proposed algorithm tends to produce a reasonably high proportion of correct singleton predictions. The average set size seems to suggest that the set-valued predictions produced by the proposed algorithms are reasonably small. Empirical evidence regarding the instability defined in (4) seems to be in favor of the proposed algorithm.

Weaknesses

W1: I think adding a few more competitors would be useful in assessing the potential advantages of the proposed algorithms. For example, threshold-based algorithms, which produce Bayes-optimal predictions of F-measure [1], and $U_{65}$ and $U_{80}$ [3], would be a good choice. They would cost $O(L\log(L))$ time to sort the labels according to the decreasing order of the predicted conditional probability masses and select the optimal threshold. W2: Assessing the classifiers with respect to utility-discounted predictive accuracies, such as $U_{65}$ and $U_{80}$ which respectively give small and large rewards for being cautious [7] , would help to further highlight the potential (dis)advantages of the classifiers. W3: Moreover, an imprecise classifier should abstain (i.e., provide set-valued predictions) on difficult cases, on which the precise classifier is likely to fail. This would be assessed in different ways. For example, one might consider reporting the correctness of the cautious classifiers in the case of abstention versus the accuracy of the precise classifiers [4]. W4: The tolerance $\epsilon$ is set to be $.05$. Please provide arguments supporting this choice. Yet, choosing this hyperparameter using a validation set might be challenging. Providing a sensitivity analysis on the choice of $\epsilon$ would be helpful. [1] Del Coz, J. J., Díez, J., & Bahamonde, A. (2009). Learning Nondeterministic Classifiers. Journal of Machine Learning Research, 10(10). [3] Mortier, T., Wydmuch, M., Dembczyński, K., Hüllermeier, E., & Waegeman, W. (2021). Efficient set-valued prediction in multi-class classification. Data Mining and Knowledge Discovery, 35(4), 1435-1469. [4] Nguyen, V. L., Destercke, S., Masson, M. H., & Hüllermeier, E. (2018, July). Reliable multi-class classification based on pairwise epistemic and aleatoric uncertainty. In 27th International Joint Conference on Artificial Intelligence (IJCAI 2018) (pp. 5089-5095). [7] Zaffalon, M., Corani, G., & Mauá, D. (2012). Evaluating credal classifiers by utility-discounted predictive accuracy. International Journal of Approximate Reasoning, 53(8), 1282-1301.

Questions

Q1: Please refer to "Weaknesses" for detailed comments and suggestions for further revision. Q2: The notion of selection stability $\delta$ at sample size $n$ seems to be interesting. I guess the experimental setting provides the necessary information to compute the value of the selection stability $\delta$ given in (1) . To provide an idea of how empirical evidence may differ from the theoretical results, it might be helpful to compute that value and report the proportion of test instances whose empirical values $1 - \delta_j$ are smaller than $\delta$. Q3: I guess re-defining (4) as $\delta_j = \frac{1}{500}\sum_{k=1}^{500} 1{s(\hat{p}(\tilde{X}_j)) \cap s(\hat{p}^{\setminus i_k}(\tilde{X}_j)) = \emptyset}$ and changing the description of Figure 2 accordingly (to be consistent with the notion of the selection stability $\delta$ defined in Definition 1) might make things a bit clearer.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

L1: Discussions on related work might need to be extended. L2: the empirical study might need to be enlarged with closely related algorithms. Please refer to "Weaknesses" for detailed comments and suggestions for further revision.

Reviewer FNaQ2024-08-09

Thanks for your detailed response. Additional empirical evidence and discussions indeed suggest more about the potential (dis)advantages of the proposed algorithms. Based on the response, I think we might see $\beta_{\text{pre}}$ as an evaluation metric for classification with rejection, i.e., a predictor will only make a prediction for a query instance if it can make a precise/singleton prediction. This might also imply that whether the set-valued prediction covers the true class or not and whether the set size is large or not are not relevant. In this sense, I think it is fair to say that the proposed algorithms provide advantages, compared to top-2, Thresholding $\Gamma^*_{80}$ and NDC for $F_1$. From their definitions, it would be not hard to see that the set-valued predictions that optimize $u_{65}$, $F_1$, $u_{80}$ and a few other criteria, e.g., ones mentioned in [6], cover the top-ranked class, i.e., the output of $argmax$ (unless there are multiple classes with the highest score). By definition, optimizing $u_{65}$ would provide smaller set-valued predictions, compared to others, and therefore might naturally become a related competitor regarding $\beta_{\text{pre}}$. Given the (set-valued) predictions on the test set, the test set can be partitioned into 2 parts, where the classifier/predictor makes precise predictions and set-valued predictions. As far as I understand from [7] (and [3]), $u_{65}$, $u_{80}$, and other utility-discounted predictive accuracies are designed to assess classifiers/predictors on the entire (test) data set. The additional results suggest that the proposed algorithms provide worse scores compared to NDC for $F_1$, which is also not designed to optimize both $u_{65}$ and $u_{80}$. If one opts for $u_{65}$ and $u_{80}$, I think it is reasonable to optimize them directly. Yet, I think it is difficult to find some specific evaluation criterion/criteria to assess multiple aspects of set-valued predictors. I think it is reasonable to expect a proposed set-valued predictor to demonstrate the advantages regarding at least a few criteria that assess the predictor on the entire (test) data set. Regarding "W3: Moreover, an imprecise classifier should abstain (i.e., provide set-valued predictions) on difficult cases, on which the precise classifier is likely to fail. This would be assessed in different ways. For example, one might consider reporting the correctness of the cautious classifiers in the case of abstention versus the accuracy of the precise classifiers [4].", I guess I should make the point more explicit. As far as I understand, the correctness of the cautious classifiers is the proportion of the time the set-valued predictions cover the true classes. I think such a criterion is designed to assess the cost of making set-valued predictions, i.e., the cautious classifiers should (only) provide set-valued predictions on difficult cases, on which the precise classifier is likely to fail. In my opinion, this criterion should complement the $\beta_{\text{pre}}$, utility-discounted predictive accuracies, set size, and other evaluation criteria, which take into account both precise predictions and set-valued predictions. Yet, I think the notion of selection stability is interesting (as I wrote in the initial review), and it might make the proposed algorithms differ from other cautious classifiers or set-valued predictors. The motivation to go with the proposed algorithms might need to be further elaborated. After due consideration, I keep my initial rating, but I appreciate any additional results and discussions that may strengthen the motivation to go with the proposed algorithms.

Authorsrebuttal2024-08-13

Thank you for your further detailed comments. We have expanded our table to include the two methods you suggest, Set-Valued Bayes-Optimal Prediction (SVBOP) for $u_{65}$ and $u_{80}$. We bold the statistical winners in each column.$^\dagger$ These new methods overall perform somewhat similar to NDC for $F_1$, but as you anticipated, the expected set size for SVBOP-65 is the smallest of the three. However, the expected set size for the inflated argmax is much smaller than these other three methods. Note also that although its $u_{65}$ utility *is* slightly smaller than SVBOP-65 on the test set, the inflated argmax qualifies as a statistical winner in this column as well. Our proposal thus performs competitively along three of the four criteria we have considered ($\beta_{\text{prec}}$, $\beta_{\text{size}}$ and $u_{65}$). We agree that, on its own, we can view $\beta_{\text{prec}}$ as an evaluation metric for classification with rejection, since it requires a singleton prediction. Our motivating goal is to return a singleton as often as possible while achieving a user-specified selection stability level. While there are many different perspectives on set-valued classification, we reiterate that what is most novel about our work is that our proposal has distribution-free theory. This means that, regardless of the dataset or base algorithm used, we can guarantee that our method will be stable. Furthermore, our optimality result shows that the inflated argmax is singleton as often as possible among all $\epsilon$-compatible selection rules. | | Results for Base Algorithm $\mathcal{A}$| | | | | Results with Subbagging $\widetilde{\mathcal{A}}_m$| | | | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | | $u_{65}$ | $u_{80}$ | $\beta_{\text{prec}}$ | $\beta_{\text{size}}$ | | $u_{65}$ | $u_{80}$ | $\beta_{\text{prec}}$ | $\beta_{\text{size}}$ | | $\text{argmax}$ | 0.879 (0.003) | 0.879 (0.003) | **0.879** (0.003) | **1.000** (0.000) | | **0.893** (0.003) | 0.893 (0.003) | **0.893** (0.003) | **1.000** (0.000) | | $\text{argmax}^\varepsilon$ | **0.881** (0.003) | 0.883 (0.003) | **0.873** (0.003) | 1.015 (0.001) | | **0.895** (0.003) | 0.897 (0.003) | **0.886** (0.003) | 1.017 (0.001) | | $\text{top-}2$ | 0.632 (0.001) | 0.777 (0.001) | 0.000 (0.000) | 2.000 (0.000) | | 0.631 (0.001) | 0.777 (0.001) | 0.000 (0.000) | 2.000 (0.000) | | Thresholding $\Gamma^*_{0.8}$ | 0.880 (0.002) | **0.909** (0.002) | 0.756 (0.004) | 1.248 (0.005) | | 0.868 (0.002) | 0.907 (0.002) | 0.703 (0.005) | 1.357 (0.006) | | NDC for $F_1$ | **0.889** (0.003) | **0.902** (0.003) | 0.832 (0.004) | 1.105 (0.003) | | **0.898** (0.003) | **0.915** (0.002) | 0.825 (0.004) | 1.137 (0.004) | | SVBOP for $u_{65}$ | **0.889** (0.003) | 0.901 (0.003) | 0.838 (0.004) | 1.091 (0.003) | | **0.900** (0.003) | **0.915** (0.002) | 0.835 (0.004) | 1.117 (0.003) | | SVBOP for $u_{80}$ | **0.885** (0.003) | **0.910** (0.002) | 0.777 (0.004) | 1.190 (0.004) | | 0.882 (0.002) | **0.915** (0.002) | 0.744 (0.004) | 1.245 (0.005) | Regarding your second suggestion, we looked at, for each imprecise classifier $\hat{s}$, the ratio of "accuracy of standard argmax given $\hat{s}$ abstains" divided by "accuracy of $\hat{s}$ given $\hat{s}$ abstains". We call this the *superfluous inflation*, since if this ratio is close to 1, it means the argmax is correct as often as the set-valued classifier (when the latter abstains), so outputting a set could be seen as overly conservative. We include the table with these results below. The inflated argmax has the smallest ratio, meaning it only abstains in hard cases where it can improve the accuracy by returning a non-singleton set. | | Results for Base Algorithm $\mathcal{A}$ | Results with Subbagging $\widetilde{\mathcal{A}}_m$ | | --- | --- | --- | | | $\beta_{\text{superfluous-inflation}}$ | $\beta_{\text{superfluous-inflation}}$ | | $\text{argmax}^\varepsilon$ | **0.496** (0.005) | **0.504** (0.005) | | $\text{top-}2$ | 0.905 (0.003) | 0.919 (0.003) | | Thresholding $\Gamma^*_{0.8}$ | 0.622 (0.005) | 0.702 (0.005) | | NDC for $F_1$ | 0.531 (0.005) | 0.591 (0.005) | | SVBOP for $u_{65}$ | 0.525 (0.005) | 0.575 (0.005) | | SVBOP for $u_{80}$ | 0.611 (0.005) | 0.689 (0.005) | $^\dagger$ To check whether a given method wins along a given column, we compute the two-sample Z-score, subtracting that method’s value from the highest value in that column and normalizing by the pooled standard error. We say method is a winner if the two-tailed Z-test is *not* statistically significant at 0.05.

Reviewer FNaQ2024-08-13

Thank you for the additional empirical evidence and discussions. To accommodate these, I increase the rating to $7$.

Reviewer mVCr4/10 · confidence 3/52024-07-13

Summary

The paper considers the stability of classifiers in multi-class classification. In the considered framework, the classifier is allowed to return a set of candidate classes instead of only one class. The stability is defined as the frequency of a set predicted by one classifier having no union with a set predicted with a classifier trained on the same training set with only one sample removed. The authors proposed the construction of a stable classifier in this framework using bagging and inflated argmax. The proposed approach is proven to have stability guarantees. Additionally, the effectiveness of the approach in terms of increased stability is confirmed in the empirical experiment.

Strengths

- The paper is nicely written and easy to follow. - The inflated argmax is shown theoretically to be the maximizer of classifier stability.

Weaknesses

- The authors argue that the argmax is a hard operator that may make the classifier unstable if even small changes of the outputs occur, but the definition of stability used in the paper also uses the hard operator of the set union being equal to an empty set. - The definition of the stable classifier as defined in this paper is new to me, and based on the paper, I struggle to find a good motivation for using this specific definition and focusing on it. - The method requires bagging, what makes it difficult to apply with heavy architectures. - The empirical results are limited to a single experiment. - The empirical experiment uses baselines that do not work in the framework of set-valued prediction and always predict only one class, which puts them at a clear disadvantage under the considered task. The comparison should include other set-valued classifiers. E.g., classifiers mentioned in related works or classifiers that optimize simple set-utility functions like in [1] and [2], that can be very efficiently optimized as discussed in [3]. [1] Del Coz JJ, Díez J, Bahamonde A. Learning nondeterministic classifiers. 2009 [2] Zaffalon M, Giorgio C, Maua DD. Evaluating credal classifiers by utility-discounted predictive accuracy. 2012 [3] Mortier T, Wydmuch M, Dembczynski K, Hullermeier E, Waegeman W. Efficient set-valued prediction in multi-class classification. 2021 NITs: - "The more recent literature has focused on producing a sparse set of weights, but none of these works offer a formal stability guarantee for the support of the weights." This statement needs citation.

Questions

- What are the specific applications benefiting from this notation of stability? - How does the introduced inflated argmax compare with other set-valued classifiers?

Rating

4

Confidence

3

Soundness

2

Presentation

2

Contribution

2

Limitations

Limitations are discussed. I see no negative social impact of this work.

Reviewer sBU32024-08-11

I thank authors detailed and patient response and I have confirmed my shared concern over limited experiments and computational costs of bagging. I thank authors for clarifying the organization of the manuscript, as well as my missing attention on section 3.2.2 and proposition 4. I also thank the improved figure 2. For my concern 1.e. for definition 1, I missed the notation of feature $x$ in $C$ and thought this considers only the classifiers sets output by an learning algorithm.

Reviewer LzW32024-08-12

Thank you for the author feedback.

Thank you very much for the author's feedback. I have read the other reviews and the rebuttal. Thank you for the explanation and additional experiments. I agree with the Author's rebuttal 2 regarding the core contribution. The core contribution of this paper is laying the foundation for stability classification by introducing several notions on classifier stability mentioned in this paper. This idea is novel to the best of my knowledge. Bagging is indeed more expensive than without Bagging, but the paper did a great job demonstrating that a theoretically justified algorithm archives this notion of stability. Having a more efficient algorithm or more analysis of this problem setting can be further studied in the future. I appreciate that other algorithms that are not necessarily stable but somewhat reasonable to try in this problem setting were added in the experiment section. This can motivate the practical usefulness of inflated argmax. Classification with abstention/rejection is indeed another framework that has been well-studied to abstain from making a prediction when the classifier is not confident. I treat this research direction as a different direction for making classifiers more reliable by outputting a flexible-size prediction set with some guarantee. This approach can also help machine learning systems to collaborate with humans to have better judgment and therefore worth investigating further. For these reasons, I increase the score to 7 (Accept).

Reviewer m1xX2024-08-13

Response to author rebuttal

I thank the authors for taking the time to addressing my concerns. I also see that some of these concerns were also discussed with other reviewers. After going through all the discussions, I think that my concerns have been addressed quite a bit. Thus I have decided to improve my score.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC