Derandomized Novelty Detection with FDR Control via Conformal E-values

Conformal inference provides a general distribution-free method to rigorously calibrate the output of any machine learning algorithm for novelty detection. While this approach has many strengths, it has the limitation of being randomized, in the sense that it may lead to different results when analyzing twice the same data, and this can hinder the interpretation of any findings. We propose to make conformal inferences more stable by leveraging suitable conformal e-values instead of p-values to quantify statistical significance. This solution allows the evidence gathered from multiple analyses of the same data to be aggregated effectively while provably controlling the false discovery rate. Further, we show that the proposed method can reduce randomness without much loss of power compared to standard conformal inference, partly thanks to an innovative way of weighting conformal e-values based on additional side information carefully extracted from the same data. Simulations with synthetic and real data confirm this solution can be effective at eliminating random noise in the inferences obtained with state-of-the-art alternative techniques, sometimes also leading to higher power.

Paper

Similar papers

Peer review

Reviewer ccpQ6/10 · confidence 4/52023-07-03

Summary

This paper applied the derandomized e-value to the conformal novelty detection, which reduces the randomness of original approach using conformal p-value. The authors also refined the method by adaptively weighting the conformal p-values based on an estimate of the out-of-sample accuracy of each underlying machine learning model. Simulations with synthetic and real data are conducted to compare the performance with the original approach.

Strengths

This paper is the first work to apply the derandomized e-value to the conformal inference, which makes it more stable compared with the original approach.

Weaknesses

### 1. The novelty of this paper is limited. Considering the previous works by applying derandomization and E-value to Knockoff filter [Ren et al., 2020, Ren and Baber 2023], it is straightforward to extend it to the conformal setting. ### 2. The challenges are not stated clearly. (1) From the connection between the conformal BH filter and Eq. (2) (given by Rava et al. [2021]), it would be easy to extend the construction in [Ren and Baber 2022] to conformal e-value in Eq. (5). In addition, the weighted e-value is also proposed by Ren and Baber [2022]. (2) Given the Theorem 2 from Ren and Baber [2022], all the technical difficulty falls in proving Theorem 3.2. However, the proof strategy is from Rava et al. [2021].

Questions

1. What is the superiority of data-driven weights over the fixed width? Is the weighted approach in Ren and Baber [2022] applicable here? 2. In Figure 2, why is the power of E-AdaDetect higher than Ada-Detect while the FDR value is still lower? It seems contradictory to common sense.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The main limitation is the novelty. Despite that this is the first work to deploy derandomized e-value to split conformal inference, the framework and the theory is well studied in previous works. In addition, the technical contribution is minor.

Reviewer TGb15/10 · confidence 2/52023-07-06

Summary

The paper employs conformal e-values, as opposed to p-values, to quantify statistical significance during outlier testing under FDR control. This approach enables the principled aggregation of results from mutually dependent tests, thereby providing a solution to de-randomize (split) conformal inferences.

Strengths

The paper addresses a significant issue of the problem of the randomness in conformal inferences. They propose a method to make conformal inferences more stable by leveraging suitable conformal e-values instead of p-values to quantify statistical significance, also merging the idea of de-randomizing conformal novelty detection and FDR control. The proposed method has the potential to significantly improve the stability and interpretability of conformal inferences.

Weaknesses

The paper could significantly benefit from revisions aimed at improving clarity. The current presentation of ideas and concepts is convoluted, making it difficult for readers to follow and understand the arguments and methodologies proposed. While the authors' approach is novel, it builds upon existing studies and techniques. The authors could strengthen the originality of their work by further highlighting the unique aspects of their approach and how it differs from previous methods. The authors provide simulations to demonstrate the effectiveness of their method, but it would be beneficial to see more empirical evaluations, including comparisons with other state-of-the-art methods.

Questions

The paper focuses on derandomizing split-conformal inferences. I wonder if jackknife+ would inherently solve the problem. The authors mention that their method can be extended to leverage adaptive weights based on the data. Could they discuss potential strategies for choosing these weights in practice?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

2 fair

Contribution

3 good

Limitations

See weakness.

Reviewer Ed417/10 · confidence 4/52023-07-07

Summary

The main limitation of conformal prediction lies in its inherent randomness. However, this paper presents an innovative solution by introducing a derandomized version of conformal prediction, specifically applied to the field of novelty detection. Through the incorporation of conformal e-values, the proposed method successfully reduces the element of randomness while providing provable and effective control over the False Discovery Rate (FDR). The key contribution of this research lies in its pioneering use of e-values, instead of traditional p-values, within the framework of conformal prediction. This innovative approach significantly simplifies the aggregation process, reducing randomness without compromising the overall detection power.

Strengths

This paper is a highly innovative and inspiring work that highlights the potential of e-values as a superior alternative to p-values for derandomizing conformal prediction through the aggregation of multiple dependent tests of the same hypothesis. The paper introduces a novel approach for constructing e-values and provides a rigorous guarantee of false discovery rate (FDR) control, with Theorem 3.2 being the main contribution of the research. Furthermore, the practical aspect of the paper lies in the deployment of conformal e-values in AdaDetect, which presents a derandomized version of AdaDetect. It is important to note that E-AdaDetect is not merely a simple combination but rather a specific practical application demonstrating how to aggregate e-values using data-adaptive weights. The experimental results showcased in the paper reveal that the derandomized AdaDetect exhibits comparable power to its randomized counterpart, while effectively controlling the FDR. This finding is particularly surprising (even higher power? )and highlights the potential benefits of adopting the "new"(not same as the classical definition) e-values .

Weaknesses

The main paper primarily focuses on experimental results obtained from synthetic data, which is sampled from simple Gaussians. It may be considered relatively easier compared to real-world scenarios. Therefore, I would suggest including at least one experiment using more complex synthetic data or real data from the supplementary material.

Questions

*

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

*

Reviewer gnkM7/10 · confidence 4/52023-07-08

Summary

This papers proposes a way to reduce the randomness in novelty detection methods (detecting out-of-distribution points) that are based on the split-conformal inference paradigm. This is done by ensembling over several ($K$) train-validation splits of the dataset. The main technical point is to aggregate the evidence from the $K$ individual predictions by averaging their E-values, as a replacement to considering the p-value in the traditional approach with only one train-validation split. It is shown that the FDR (false discovery rate) can still be controlled with these quantities. The method is further enhanced by using certain weighted averages of the E-values, taking into account the estimated powers of each one. The methods are evaluated on synthetic and real data, showing that the proposed derandomized method works better if the fraction of outliers is larger.

Strengths

The proposed novelty detection approach is original in that it combines evidences from different algorithm runs into a single evidence score. This required a formulation of the algorithms' evidences in terms of E-values instead of p-values that are commonplace for conformal methods, to avoid a loss in detection power. While prior novelty detection works have touched upon E-values, the current paper succeeds in controlling the FDR based on these E-value evidences. The methodological details (Sec. 3.1) and mathematical proof (Theorems 3.2 and 3.6) of this FDR control are strongly inspired by existing works, but this technique is new in the area of novelty detection. As to significance, the method does reduce the variance as promised and keeps its FDR guarantee. It has higher power than competitor methods in some regimes, but quite consistently weaker power in other regimes, particularly for low fraction of outliers. Secondly, while the problem and the solution is built around quite specific requirements, it is plausible that the presented techniques can be applied to the derandomization of other inference methods with statistical guarantees, as the authors say in their Conclusion. The clarity of the paper is exceptional (but see my comments on Appendix S2 below), the explanations and discussions are to the point.

Weaknesses

As the biggest weakness of the paper I see that, despite the novelty of the approach, the considered problem requirements are relatively specific and the solution therefore narrow, as it concerns the de-randomization of certain novelty detection algorithms that are based on conformal inference and that aim at a mathematical control of the FDR. Also, the method's performance (power) falls behind other methods in certain application regimes. It would be good to have some rules beforehand to know in which regime to apply which method. If space permits, I would like to see some real data experiments (see Sec. 4.3) in the main text rather than only in the Supplement.

Questions

* Please mention what K is in Fig. 1. * line 228: Should this be D^(k)_cal \cup D_test? * How is AdaDetect different from E-AdaDetect with K=1 (Fig. 3)? * Please carefully re-write the Proof of Theorem 3.6 in Appendix S2, as I believe there are many typos and several things that could be explained better. In particular: - Please explain the first equality after line 35. - It seems that the martingale runs from l'=l down to l'=0, correct? - It would be good to explain in detail, over which random variables the various expectations E after line 26 run and which random variables are held fixed (conditioned on). - Right-hand-sides of Eq. (S1) and (S2): Should there be t instead of \hat{t}^{(k)}? - line 31: The word "outliers" should be "inliers" - Should the inf be changed to sup, or maybe the inequality sign reversed? - line 34: inside the curly brackets, the l should be l'?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

Yes, this is discussed in a fair way after line 351.

Reviewer ccpQ2023-08-16

Response to rebuttal

Thank you for answering my questions. As you said, data-adaptive weighting method is the main innovation of this paper. However, the corresponding results are not sufficient. I think the technical difficulty of data-driven weights is addressed by the symmetric assumption (line 206). Can you extend the results to more general cases? Also, there is no experiments on real data to show the superiority of data-driven weights.

Authorsrebuttal2023-08-17

Re: results of additional experiments based on real data

Thank you for the suggestion of including additional experiments to demonstrate the performance of our proposed data-driven weighting method on real data. Following your suggestion, we have carried out additional experiments using the same 4 real data sets considered in the paper: "musk", "shuttle", "KDDCup99", and "creditcard". We have uploaded the results, along with a detailed description of the setup, in this one-page blinded PDF file: https://docdro.id/8QEqtk8 In summary, the results demonstrate clearly the advantage of the two alternative weighting schemes ("t-test" or "avg. score"), which in most cases lead to noticeably higher power compared to the benchmark with uniform weighting. Notably, we see that weighting based on the t-test is the most powerful approach. These results are consistent with those presented in Figure 4 of the paper, which were based on synthetic data. Further, it is interesting to note from Figure 1 (in the new one-page PDF) that data-driven weighting is also effective at further reducing the algorithmic randomness of our findings, leading to lower variance. We would of course be happy to include these results in the revised manuscript, possibly along with other similar results which we did not include in the one-page PDF for brevity. Finally, regarding your suggestion of extending our data-driven weighting method to more general cases (e.g., classification or regression, instead of outlier detection), this is certainly a good idea for future work. However, we feel that it would go beyond the scope of this paper to investigate methods for problems other than outlier detection, partly also due to to space limitations.

Reviewer gnkM2023-08-18

Thank you for addressing my questions. I appreciate that you highlight the exchangeability as the central part of conformal inference, and that this comes out particularly in novelty detection. I am looking forward to seeing your approach being extended to derandomize other conformal tasks. Thanks also for your insights on what to expect on power. With this in mind, it is even surprising that your derandomized method improves power consistently in some regimes. I still encourage you to take some real-data experiments into the main paper, potentially even some of the ones you prepared as a response to Reviewer ccpQ during the rebuttal. I am raising my score to 7 and am recommending acceptance of the paper.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC