Overcoming Common Flaws in the Evaluation of Selective Classification Systems

Selective Classification, wherein models can reject low-confidence predictions, promises reliable translation of machine-learning based classification systems to real-world scenarios such as clinical diagnostics. While current evaluation of these systems typically assumes fixed working points based on pre-defined rejection thresholds, methodological progress requires benchmarking the general performance of systems akin to the $\mathrm{AUROC}$ in standard classification. In this work, we define 5 requirements for multi-threshold metrics in selective classification regarding task alignment, interpretability, and flexibility, and show how current approaches fail to meet them. We propose the Area under the Generalized Risk Coverage curve ($\mathrm{AUGRC}$), which meets all requirements and can be directly interpreted as the average risk of undetected failures. We empirically demonstrate the relevance of $\mathrm{AUGRC}$ on a comprehensive benchmark spanning 6 data sets and 13 confidence scoring functions. We find that the proposed metric substantially changes metric rankings on 5 out of the 6 data sets.

Paper

References (57)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer QM337/10 · confidence 3/52024-06-26

Summary

The authors introduce a new metric AUGRC for evaluating classifiers under the Selective Classification framework, whereas the classifier has an option to reject low-confidence predictions. The authors introduce desirable properties for evaluating the Selection Classifiers, and show that the new metric has all those properties, unlike any of the currently used metrics. Experiments show that AUGRC produces significanly different rankings of Confidence Scoring Functions compared with currently used metrics.

Strengths

The authors address relevant question and make useful contribution. The paper is well written, clear and easy to read. The experimental analyses are, to my judgement, sound. The code is provided.

Weaknesses

Fig. 4: the risk vs. coverage curves for AURC vs. AUGRC do not look that different. Sure, there is non-monotonicity in AURC, but that is in the low-coverage region, which presumably is not of practical interest. The rest looks monotonic and fairly similar. It is also true that the two metrics suggest different CSFs, but again, they aren't that different. To be clear, I still think AUGRC is favorable, just saying that the difference appears modest. I think the key conclusion, which is the recommendation to adopt AUGRC, is valid, however the language in the section is a bit too strong (there is "substantial" twice and "significant limitations"). I recommend to tone-it-down.

Questions

The authors claim that Selective Classification is a crucial components for clinical diagnostics ML systems (among other application domains). Can you provide examples of classifiers in clinical use that incorporate Selective Classification? What is AUROC_f? It is introduced on line 172, however the description is unclear and there are no references.

Rating

7

Confidence

3

Soundness

3

Presentation

4

Contribution

3

Limitations

There is no limitations section. I don't have a good idea regarding limitations of this manuscript.

Reviewer izgv6/10 · confidence 4/52024-06-27

Summary

The paper presents 5 requirements for multi-threshold metric for Selective Classification and a novel metric to evaluate selective classifiers called AUGRC. The proposed metric satisfies 5 requirements that are not met by current approaches. The proposed metric changes the rankings on 5 out of 6 datasets considered by the authors.

Strengths

The paper's main strengths are: 1. the paper aims to tackle a long-standing concern in abstaining classifiers literature, i.e. how to evaluate with a single measure these classifiers 2. the theoretical derivations seem sound 3. the contribution is well framed within current literature

Weaknesses

The main concerns of the paper are: 1. The empirical evaluation can be improved 2. The interpretability of the proposed measure is not that straightforward 3. The paper presentation can be improved. Regarding the empirical evaluation, I have a few remarks. * there are some contradictory lines, which should be double-checked: for instance, in lines 268-269, it is not clear to me why you claim that AURC erroneously favors DG-RES over MCD-PE, as DG-RES's performance seems to be better than MCD-PE (accuracy-wise and ranking-wise). Similarly, it is not clear to me why in Figure 4 the authors state DG-RES is favored "despite a lower classification performance and ranking quality", while in Figure 4.a is reported a higher accuracy and higher $AUROC_f$ for DG-RES; * second, the authors never specify whether they correct for multiple outcomes testing (e.g. using Bonferroni correction). I think since the authors are performing multiple pairwise ranking tests, they also have to account for this. Regarding the interpretability requirement, the authors claim (correctly) that the AUROC can be interpreted as "the probability of a positive sample having a higher score than a negative one". I did not fully grasp what is the straigthforward interpretation of the AUGRC score in this context (lines 206-208). Regarding paper presentation, the paper heavily relies on acronyms. I would personally reconsider this choice to improve the overall readability.

Questions

I have a few questions for the authors: * [Q1] can the authors clarify my concerns regarding the empirical evaluation? * [Q2] can the authors discuss further my concern regarding interpretability? Is there a straightforward interpretation of the AUGRC score? * [Q3] when considering a degenerate abstaining classifier that always abstains, i.e., whose confidence score is always zero, is the AUGRC 0 (as the one obtained by a perfect classifier)? If so, is AUGRC favoring classifiers that tend to be underconfident and asbtain too much? Is there a way to avoid such a behaviour? * [Q4] if I correctly understand how AUGRC works, why do not the authors consider the _classification quality_ by [Condessa et al, 2017] instead of the Generalized Risk Score? Such a metric also considers the failures that are correctly rejected and could avoid favoring abstaining classifiers that over reject. Typos: line 128: schore $\to$ score

Rating

6

Confidence

4

Soundness

2

Presentation

2

Contribution

3

Limitations

The authors do not directly discuss limitations of their proposed approach in the main paper. For instance, I think a brief discussion regarding the computational time required to compute AURGC should be included in the main paper.

Reviewer xFt78/10 · confidence 5/52024-07-06

Summary

The authors propose 5 requirements that should be satisfied by selective classification (SC) metrics such that they can be successfully used to rank SC models for a task. They then propose a new metric, called AUGRC, which is shown to satisfy all 5 requirements. Finally, the authors show empirically that their method performs better rankings thank previous metrics.

Strengths

1. The paper is well-motivated, clearly written, and has an excellent coverage of previous works. 2. The 5 requirements were well-chosen and a metric that satisfies them is likely to be a good choice for SC ranking. 3. The metric is simple, yet effective, and the maths seems sound. 4. Results are convincing and statistically analysed; the toy dataset was didactic. 5. There's enough information in the main body of the paper to fully understand the experiments and corresponding results, i.e. it's not necessary to read the appendix, even if further details are available there.

Weaknesses

1. Limited discussion of future works. 2. NLL and Brier Score are listed among the multi-threshold metrics, but there's no thresholding involved in their evaluation. In fact, it should be possible to use them as risk measures, which would cover more for than just 0/1 loss, especially where probabilities are important for decision making.

Questions

1. A similar task to SC is classification with rejection of OOD samples (sometimes called open-set recognition). In such cases, there isn't a ground truth to evaluate the risk on (in practice). Can the authors envision a way to adapt their metric to these scenarios? 2. The metrics were scaled by ×1000 in Figure 2. Were they also rescaled for all the other results? 3. Regarding interpretability, the authors say that "The optimal AUGRC is given by the second term in Equation 7". What does the optimal AUGRC mean? What would be the AUGRC of the Bayes-optimal model for a task? Is there an AUGRC value associated to a random classifier, as in AUROC? 5. Any intuitions about why the rankings of AURC and AUGRC where the same for iWildCam?

Rating

8

Confidence

5

Soundness

4

Presentation

4

Contribution

4

Limitations

Yes.

Reviewer BBe87/10 · confidence 4/52024-07-16

Summary

The paper tackles the problem of Selective Classification (SC). The authors show problem with the existing metrics used in evaluating SC and propose to use the Area under the Generalized Risk Coverage curve (AUGRC). Empirical results provide useful insights about the effect of using this metric and shows how this changes the relative ordering of state-of-the art methods.

Strengths

- The problem tacked is very important, especially for the safe deployment of ML models which requires the knowledge of when not to trust the model. - The proposed metric is well-motivated. - The experiments are extensive. - It is important for the community to know about this work as this changes the perception of SC and the best method to use.

Weaknesses

- It is important to add the details of the methods in the main paper. For example, DeepGamblers (DG) is referred to as DG in the main paper and this abbreviation is only explained in the appendix. Similarly for other baselines. - Although the experiments are extensive, it is good to include the recent state-of the art SC methods such as Feng et al (2023) to understand which method is the best to use. References: Feng et al: Towards Better Selective Classification

Questions

See the previous section

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

NA

Reviewer izgv2024-08-08

Response to the Rebuttal

I thank the authors for their clarifications and the new experiments. A few comments: regarding Q1, I thank the reviewers for implementing my suggested changes. regarding Q2, > the interpretation is: On drawing two random samples, it corresponds to the probability that either both are failure cases or that one is a failure and has a higher confidence score than the non-failure. This can be read from Equation 7: I think clarifying the interpretation of AUGRC (meaning adding these lines to the paper) is very important and can help the reader improving its understanding. Regarding Q3 and Q4, I see your points, and I have no further questions. Finally, I thank the authors for providing a small analysis regarding the required time for computing AUGRC. Hence, after the authors' rebuttal and clarifications, I will increase my score to a weak accept.

Authorsrebuttal2024-08-13

Thank you for your response

Thank you very much for taking the time to read our response and re-considering your assessment based on the provided updates.

Area Chair xygR2024-08-12

Author-Reviewer Discussion Phase

Dear Authors and Reviewers! Thank you for your reviews, rebuttal, additional comments, questions and responses! We have the last two days of the discussion phase! Please use this time as efficiently as possible :) Thank you, NeurIPS 2024 Area Chair

Authorsrebuttal2024-08-13

Thank you for your response

Thank you very much for taking the time to read our response and re-considering your assessment based on the provided updates.

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC