Statistical Multicriteria Benchmarking via the GSD-Front

Given the vast number of classifiers that have been (and continue to be) proposed, reliable methods for comparing them are becoming increasingly important. The desire for reliability is broken down into three main aspects: (1) Comparisons should allow for different quality metrics simultaneously. (2) Comparisons should take into account the statistical uncertainty induced by the choice of benchmark suite. (3) The robustness of the comparisons under small deviations in the underlying assumptions should be verifiable. To address (1), we propose to compare classifiers using a generalized stochastic dominance ordering (GSD) and present the GSD-front as an information-efficient alternative to the classical Pareto-front. For (2), we propose a consistent statistical estimator for the GSD-front and construct a statistical test for whether a (potentially new) classifier lies in the GSD-front of a set of state-of-the-art classifiers. For (3), we relax our proposed test using techniques from robust statistics and imprecise probabilities. We illustrate our concepts on the benchmark suite PMLB and on the platform OpenML.

Paper

Similar papers

Peer review

Reviewer n7Tx7/10 · confidence 4/52024-07-06

Summary

The authors propose a novel way of comparing classifiers to assess their effectiveness. They posit (1) that comparisons should allow for different quality metrics simultaneously, (2) that comparisons should take into account the statistical uncertainty induced by the choice of benchmark suite, and (3) that the robustness of the comparisons under small deviations in the underlying assumptions should be verifiable. They illustrate their proposed method on the benchmark suite PMLB and on the platform OpenML.

Strengths

The paper is clear and easy to follow. The proposed method is innovative, it enjoys some desirable mathematical properties (that are well-presented), and it seems to work well on the selected benchmarks.

Weaknesses

Some references to the existing imprecise probabilistic literature are missing, there are some small typos, and an extra effort can be put to make the definition of $d_s$ more accessible.

Questions

When introducing imprecise probabilities, the authors fail to reference a few papers in such a literature. In particular, the authors leverage the idea of $\epsilon$-contamination in section 4.2. Other papers in the imprecise probabilistic machine learning literature dealing with, and advancing the knowledge of, $\epsilon$-contaminations are [1-5]. In addition, [6] uses IPs to robustify Bayesian neural networks. More in general, we suggest the authors to read (and possibly cite) the works by Eyke Hüllermeier, Thierry Denoeux, Sebastien Destercke, Fabio Cuzzolin, Alessio Benavoli, and Michele Caprio on imprecise probabilistic machine learning, all of which have a distinct "robustness" flavor to them, and that might be interesting to the authors for future endeavors. There is a typo in line 95: there's a period after superscript 3. While I understand the need of being as general as possible, wouldn't it be more beneficial to the reader to define $\hat{\pi}_c(\{z\})= 1/s \left|\lbrace\{ i : i \leq s \text{, } \Phi(C,T_i)=z \rbrace\}\right|$ in line 167, instead of using a generic $M \subseteq [0,1]^n$ for the definition? There is a typo in line 167: last letter should be $C^\prime$ and not $C$. Line 205: the question *of* how to [...] --- [1] https://www.jstor.org/stable/2242055 [2] https://onlinelibrary.wiley.com/doi/book/10.1002/9780470434697 (Chapter 10) [3] https://link.springer.com/chapter/10.1007/978-3-031-57963-9_1#:~:text=In%20their%20seminal%201990%20paper,bound%20to%20hold%20with%20equality. [4] https://arxiv.org/abs/2402.00957 [5] https://arxiv.org/abs/2308.14815 [6] https://arxiv.org/abs/2302.09656

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

See questions.

Reviewer 3sEe8/10 · confidence 4/52024-07-09

Summary

The authors propose a method for multicriteria evaluation of classifiers which is more informative than the Pareto front. The new GSD-front and is based on the previously proposed generalized stochastic dominance ordering (GSD) for classifiers. The authors also provided a sound inference framework, included hypothesis tests for whether a new classifier would be in the GSD-front of a benchmark.

Strengths

1. The method is sound and extremely useful 2. The paper is very well written; notation is consistent throughout 3. Given the large (and increasing) number of potential models that can be applied to any given task, this method is very significant for identifying which models tended to be among the best for other similar tasks 4. The hypothesis tests derived from the framework are also very useful to add significance to these benchmarks 5. All the information needed to fully understand the paper is in the main body of text

Weaknesses

1. There could have been maybe a toy example to show the differences between GSD-front and Pareto-front in a more didactic way 2. The method seems computationally expensive. The pmlb_permutation_tests.R script needs 5 days to run and this does not include training of the evaluated models, as these are collected results

Questions

1. How does the running time of a Pareto-front benchmark compare with a GSD-front one? 1. If the Pareto-front is much cheaper to obtain, would it make sense to start with it, and only look for what classifiers in that pre-selection would be in the GSD-front? 2. How were different hyper-parameter values treated in the experiments? Were all results of each classifier for a dataset aggregated or did the authors use specific settings for each classifier, e.g. default values from the classifier's implementation?

Rating

8

Confidence

4

Soundness

4

Presentation

3

Contribution

3

Limitations

Yes

Reviewer 1fNM7/10 · confidence 3/52024-07-12

Summary

This submission studies the problem of comparing multiple classifiers under multiple evaluation criteria. It presents the construction of the GSD-front, i.e., the set of GSD-undominated classifiers, and an empirical variant called the $\epsilon$-empirical GSD-front. Then, theoretical aspects of the GSD-front and the $\epsilon$-empirical GSD-front, such as satistical consistency, and the inclusion concerning the respective Pareto-front, are investigated. Two statistical tests, namely static GSD-test and dynamic static GSD-test, on whether a given classifier $C \in \mathcal{C}$ is an element of the true GSD-front are proposed. Both tests are shown to be valid level-$\alpha$ tests. A robustified static GSD-test is also proposed to deal with the setting where the available data sets are drawn under non-i.i.d.-scenarios. An empirical study is conducted to assess the proposed tests. It covers 80 binary classification data sets from OpenML [70]. The set of classifiers consists of Support Vector Machine (SVM) with Random Forest (RF), Decision Tree (CART), Logistic Regression (LR), Generalized Linear Model with Elastic net (GLMNet), Extreme Gradient Boosting (xGBoost), and k-Nearest Neighbors (kNN). The set of evaluation metrics consists of predictive accuracy, computation time on the test data, and computation time on the training data. Empirical evidence suggests that the GSD-approach tends to be more contrastive, compared to Pareto-front and another test that aggregates the single-metric tests as in [19]. Another empirical study is conducted using 62 datasets from the Penn Machine Learning Benchmark (PMLB). The set of classifiers consists of rule ensembles of trees (CRE) [53], CART, RF, SVM with radial kernel, kNN and GLMNet. The set of evaluation metrics consists of classical accuracy (metric), ii) robustness of accuracy w.r.t. noisy features (ordinal), and iii) robustness of accuracy w.r.t. noisy classes (ordinal). The results provided by GSD-tests are discussed.

Strengths

S1: The submission is well-written and organized. S2: Theoretical aspects of the GSD-front and GSD-tests are carefully investigated. S3: The authors also take into account an interesting aspect that the different evaluation metrics may have different levels of impact. Non-i.i.d.-scenarios are also taken into account. S4: Empirical evidence seems to suggest that GSD-tests can complement the existing statistical tests for comparing classifiers meaningfully.

Weaknesses

W1: The authors might consider mentioning more scenarios where comparing multiple classifiers under the multiple evaluation criteria is crucial. An example might be multi-label classification, where multiple evaluation metrics have been proposed to evaluate multi-label classifiers. The set of evaluation metrics can also be extended with others, such as training time, storage memory, and so on. W2: The motivation to compare multiple classifiers using evidence on multiple data sets, which can come from very different application domains might need to be strengthened. For example, it would be not easy to convince practitioners in safety-critical applications, such as lung cancer detection to use some classifier that has shown to be promising given evidence aggregated from multiple domains. W3: Additional recommendations for the end-user might be beneficial. For example, what should one do if the GSD-front contains multiple classifiers?

Questions

Q1: Could you elaborate more on the case where only one training data is taken into account? This is related to the comment in W2. Q2: Regarding W3, should one randomly choose one element of the GSD-front when having to make predictions? Yet, one might consider doing ensemble learning on the set of promising classifiers. In safety-critical applications, such a strategy might lower the interpretability and explainability of the predictive system. Do you have any recommendations in such cases? Q3: Could you also discuss the Pareto-front and results of the test that aggregates the single-metric tests as in [19] for the empirical study on PMLB data sets?

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Please refer to "Weaknesses" and "Questions" for detailed comments.

Reviewer 1fNM2024-08-09

Thanks for your reply. I think the response indeed complements the submission. The additional discussion on the case where GSD-front contains multiple classifiers might need to be further elaborated as it might be of practical relevance. In different applications of credal decision making, I guess one might expect that there are experts who know and can provide the true classes (by possibly paying additional costs). Yet, I think whether one should expect such experts in model selection/comparison might require more careful attention. However, I think the motivation for avoiding overoptimism is also interesting. After due consideration, I raised the rating to $7$.

Authorsrebuttal2024-08-09

Thanks for the reply

Thank you for your reply and for taking into account our rebuttal in your assessment of the paper! We will include a discussion about the case where the GSD front contains multiple classifiers in the new paragraph with end-user recommendations.

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC