acceptodds
Under review as a conference paper at ICLR 2027

Generalizing Fair Top- Selection: An Integrative Approach

Abstract

We study the problem of finding a fair (linear) scoring function with multiple protected groups (i.e., minority or disadvantaged groups) while also minimizing the disparity from a reference scoring function, generalizing the prior setup that was restricted to the single-group setting without disparity minimization. Driven by the need for experimental exploration, we find that previous studies overlook a critical issue that may affect the fairness of the outcome. Once it is properly considered, the problem may become computationally intractable even for a two-dimensional dataset and small . However, our analysis also reveals a gap that enables us to recover the efficiency for small when the number of protected groups is sufficiently small. Furthermore, we introduce a new disparity measure—utility loss—that may yield a more stable scoring function under small weight perturbations. Through careful engineering trade-offs that balance implementation complexity, robustness, and performance, our augmented two-pronged solution demonstrates strong empirical performance on real-world datasets, with experimental observations also informing algorithm design and implementation decisions.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Fairness, Top-k, Integration

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_KNjTOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies fair top-(k) selection under constraints on multiple protected groups, minimizing either the (L_1) distance from a reference scoring vector or utility loss. Its contributions include hardness and tractability results, extensions of (k)-level algorithms to handle ties and multiple groups, and a MILP-based approach for larger (k). Experiments on COMPAS and IIT-JEE demonstrate substantial runtime improvements over baseline algorithms.

Critical strengths
  1. Explicitly handles ties and multiple protected groups, which can materially affect fairness feasibility.
  2. Connects complexity analysis, geometric algorithms, and MILP, with substantial reported runtime improvements in some experiments.
Critical weaknesses
  1. The multidimensional baseline does not handle ties, so it is difficult to tell how the proposed methods would compare against a baseline with the same capabilities in terms of runtime and fairness guarantees.
  2. The stability evaluation covers only ten scoring vectors, with four perturbations per vector at a single magnitude. This makes it difficult to determine whether the observed advantage holds under other conditions.
Questions for authors
  1. How would the proposed methods compare with baselines that also handle ties and provide the same fairness guarantees?
  2. How were the stability perturbations generated, and would the results hold under different perturbation settings?
  3. What are the main theoretical and algorithmic contributions that go beyond the existing (k)-level-based methods cited in the paper?
3 comments

We thank the reviewer for their comments and questions. Please find our answers below.

W1. The multidimensional baseline does not handle ties, so it is difficult to tell how the proposed methods would compare against a baseline with the same capabilities in terms of runtime and fairness guarantees.

Q1. How would the proposed methods compare with baselines that also handle ties and provide the same fairness guarantees?

In Appendix H.1 (line 1457), it has been stated that "the tie-breaking was ignored in the implementation of ATC+ (the multidimensional baseline), as it was not obvious how to do so without introducing significant runtime overhead for ATC+". Making ATC+ correctly handles ties requires additional computational steps, increasing its run time. This means our performances advantages will be even greater when compared with the corrected ATC+ variant. Given that our experiments have already demonstrated speedups of several orders of magnitude, we believe these results sufficiently support our runtime efficiency claim over ATC+ (the multidimensional baseline). Furthermore, given ATC+'s clear runtime inefficiency relative to our method, we use it solely to demonstrate the runtime advantage and do not consider it worthwhile to develop a corrected ATC+ variant.

W2. The stability evaluation covers only ten scoring vectors, with four perturbations per vector at a single magnitude. This makes it difficult to determine whether the observed advantage holds under other conditions.

As stated in Section 6.2 (lines 419–420): "These (stability evaluation) results suggest that optimizing for the utility loss can yield more stable fair weight vectors than optimizing for a distance-based objective, aligning with our theoretical analysis (Section 4.2)". Our claim regarding the stability advantage is primarily supported by our theoretical analysis rather than empirical observations. Specifically, in Section 4.2, we explain that: "A fair weight vector minimizing the w difference (i.e., L1 distance) is inherently unstable, as a small perturbation towards wow^o would yield top-k subsets violating the fairness constraints." (lines 264–265). In contrast, for a fair weight vector minimizing the utility loss, we state: "This linear program enforces a margin of ξ between the weight vector and the cell boundary, ensuring that a weight vector under small perturbations bounded by ξ induces the same top-k subset (with the minimum utility loss). As a result, the resulting fair weight vector is more stable with respect to perturbations (see Figure 2)" (lines 270–274). The intention of our stability evaluation experiment is only to empirically demonstrate the stability advantage of the utility loss objective over the distance-based objective. The stability advantage, by itself, is established via a theoretical analysis. We will clarify the intended role of this experiment in the revised version.

Of course, as stated in the Reproducibility Statement, the command-line interface (CLI) tool provided in our (anonymous) code repository "supports experimental configurations beyond those reported in the paper" (Lines 498–499). Here, we report additional experimental results that for 25 weight vectors, with 8 perturbations per vector at a perturbation L1 distance of 2×10−42\times 10^{-4}, 24 of 25 fair weight vectors found under the utility loss objective (with stabilization) passed the stability test. In contrast, all fair weight vectors found under the w difference objective failed. However, we believe our original experimental results (10 vs. 0) are good enough to empirically demonstrate the stability advantage, as the claim is primarily supported by our theoretical analysis rather than empirical observations.

Q2. How were the stability perturbations generated, and would the results hold under different perturbation settings?

We have briefly mentioned how were the stability perturbations generated in Section 6.2 (lines 414–416): "For each fair weight vector, four randomly perturbed weight vectors were generated with an L1 distance of 10−410^{-4} from it". That is, each perturbed weight vector is obtained by randomly perturbing the weights of the original fair weight vector, but subject to two constraints: all weights in the perturbed weight vector are non-negative and sum to 1 (i.e., its L1 norm is 1), and its L1 distance from the original fair weight vector (i.e. the sum of the absolute differences between corresponding components of the weight vectors) is exactly 10−410^{-4}. We will clarify this in the revised version. While one can certainly test other different perturbation settings (see our reply to W2 for additional results), the stability advantage is primarily established through our theoretical analysis rather than empirical observations. Therefore, we believe that the current experimental results are good enough for our purposes.

Q3. What are the main theoretical and algorithmic contributions that go beyond the existing (k)-level-based methods cited in the paper?

For theoretical contributions, we prove the Fair Top-k Selection problem is NP-hard with multiple protected groups once the tie-breaking issue is properly considered, thereby correcting an implication of previous work. Furthermore, we establish a near Ω(nk)\Omega(n^k) lower bound for constant values of k under a well-known conjecture in fine-grained complexity. For algorithmic contributions, guided by the hardness analysis, we propose an efficient algorithm for the Fair Top-k Verification problem when both the number of protected groups and k are small. This algorithm is a key component to augment the k-level based algorithm to recover its (theoretical) efficiency for small k, while also constituting an algorithmic contribution in its own right through its important role in experimental exploration. Moreover, we present an augmented MILP-based algorithm for large k.

Finally, as we have argued throughout our submission, we acknowledge that some of these contributions may be of independent interest, still, we emphasize that their primary significance in this work lies in their interconnection, rather than their individual contributions considered in isolation (see Appendix A for further discussion).

Thanks for the response. My concerns are addressed. I will keep my score.

Reviewer_apbyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studied a fair top-k selection problem among overlapping protected groups. With linear scoring rules, the authors proposed to select fair weights based on utility loss. They showed ties make fair top-k verification NP-hard even in two dimensions. They used k-level and mixed-integer linear programming (MILP) algorithms to find the optimal solution by adding utility preserving stabilization. They showed the proposed algorithm outperformed k-level and MILP based on their numerical results on COMPAS and IIT-JEE datasets.

Critical strengths

Their reported numerical results based on COMPAS and IIT-JEE datasets validated the performance of their proposed optimization algorithm.

Critical weaknesses
  1. In Section 4.2 and equation (5) , the authors stated that minimizing the utility loss objective yields a stable fair weight vector. However, an optimal fair selection may exist only on a tie boundary. An explicit positive stability margin condition is needed.

  2. The limited numerical studies do not provide sufficient evidence to support their conclusions regarding the proposed fair top-k selection method.

Questions for authors
  1. Under what conditions is the optimum margin in equation (5) strictly positive?

  2. The limited numerical studies do not provide sufficient evidence to support their conclusions regarding the proposed fair top-k selection method. Can the authors conduct extensive simulation studies to evaluate the performance of their fair top-k selection method across different settings?

3 comments

We thank the reviewer for their comments and questions. Please find our answers below.

W1. In Section 4.2 and equation (5), the authors stated that minimizing the utility loss objective yields a stable fair weight vector. However, an optimal fair selection may exist only on a tie boundary. An explicit positive stability margin condition is needed.

Q1. Under what conditions is the optimum margin in equation (5) strictly positive?

We agree that in certain situations, one may only find the fair weight vector minimizing the utility loss on the cell boundary. However, when the top-k subset induced by the fair weight vector minimizing the utility loss corresponds to a (k−1)(k-1)-level cell, the optimum margin in equation (5) must be strictly positive. In particular, this holds when such a top-k subset has no candidates tied at the selection boundary (i.e., no other candidate has the same score as the kth one), or when all candidates tied at the boundary have identical values across all dd scoring attributes. Therefore, we believe that the stabilization works in many practical situations. We will add this discussion to the revised version.

W2. The limited numerical studies do not provide sufficient evidence to support their conclusions regarding the proposed fair top-k selection method.

Q2. The limited numerical studies do not provide sufficient evidence to support their conclusions regarding the proposed fair top-k selection method. Can the authors conduct extensive simulation studies to evaluate the performance of their fair top-k selection method across different settings?

Due to the page limit, we report only representative runtime experimental results in the main body of the paper, which are intended to support our key experimental observations. More extensive results covering a wide range of configurations, together with discussions of these results, are provided in Appendix H.2. Furthermore, as stated in the Reproducibility Statement, the command-line interface (CLI) tool provided in our (anonymous) code repository "supports experimental configurations beyond those reported in the paper" (lines 498–499). Nevertheless, we believe that the runtime improvements already demonstrated are substantial enough to convincingly support our performance claims, achieving up to a 50× speedup in 2-D experiments and speedups of several orders of magnitude in multi-dimensional experiments over the baseline. As an additional note, we do not use synthetic datasets in our experiments as explained in Appendix H.1 (lines 1446–1449), following (Cai SoCG'26).

Reference

Guangya Cai. Finding a fair scoring function for top-k selection: From hardness to practice. SoCG 2026.

Thank the authors for their effort and responses. I have decided to keep my score unchanged since the novelty of this work remains insufficient relative to Cai's work.

We thank the reviewer's eventual reply.

While the reviewer stated that they decided to keep their score unchanged "since the novelty of this work remains insufficient relative to Cai's work", they did not raise any concerns regarding the novelt of our work in their original comments. So, we also do not see any sufficient justification for this judgment by this single statement in the follow-up comment alone. In any case, we would like to reiterate the conclusion of our submission, which emphasizes the main contribution of our work that we have consistently argued throughout the paper (Introduction, Figure 1, Conclusion, Appendix A, etc.).

We presented an integrated study of the fair top-k selection problem in a generalized setting. While one may read our work by extracting the locally defined contribution of each individual component and ignoring its nonlinear structure (Figure 1), we believe the true crux of this work lies in how these components are interconnected. It is these connections that lead to new insights into the problem and enable our superior solution. Ultimately, the whole is greater than the sum of its parts.

Reviewer_44QLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work studies fair top-k selection by learning a linear scoring function that satisfies proportional representation constraints for multiple, potentially overlapping protected groups while minimizing deviation from a reference scoring function. The authors extend prior work which was restricted to a single protected group. They investigate the computational challenges introduced by multiple groups, particularly when candidates have tied scores. They establish NP-hardness for fair top-k verification even in two dimensions and a conditional fine-grained lower bound for small k. They then exploit the limited number of group-membership profiles to develop an efficient verification algorithm when the number of protected groups is constant and k is sufficiently small.

Building on these theoretical results, the paper extends existing k-level-based and mixed-integer linear programming algorithms to handle multiple protected groups and two disparity objectives: weight difference and utility loss. They also facilitate selecting scoring functions that are more robust to small perturbations. Experiments on COMPAS and IIT-JEE demonstrate improved optimization objectives and substantial runtime advantages over existing baselines, with additional experiments examining scoring-function stability.

Critical strengths
  1. In theory, the paper identifies a real correctness gap in earlier multi-group methods that ignore ties, which can change whether a weight vector is fair. The NP-hardness and conditional lower-bound results establish meaningful limitations on efficient fair top-k verification. More importantly, the authors identify a tractable regime by exploiting the exchangeability of candidates sharing the same group-membership profile. This provides a principled basis for the proposed efficient algorithms.

  2. In experiments, the paper translates the theoretical observations into practical extensions of k-level-based and MILP-based algorithms, supporting both multiple protected groups and disparity minimization. The experiments show that their solution is substantially faster than the baselines. The utility-loss objective provides an additional mechanism for obtaining more stable scoring functions.

Critical weaknesses
  1. Unclear practical usage and limited novelty relative to Cai.

The paper considers a scoring function fair whenever at least one of its possible top-k subsets satisfies the fairness constraints. However, in real-world selection systems, tie-breaking is typically determined by a fixed or randomized procedure. Consequently, the existence of a fair tie-breaking outcome does not ensure that the deployed selection process will produce a fair result. Although the paper explicitly addresses ties algorithmically, the distinction between existential feasibility and operational fairness warrants more careful consideration.

Additionally, the algorithmic contributions are largely incremental augmentations of the prior paper’s two-pronged solution, and Theorem 4’s runtime is inherited plus polylog overhead. The headline hardness results come entirely from fully degenerate instances where every candidate ties. That is a combinatorial covering problem, and its hardness is unsurprising and largely known. The “even in 2-D” framing and the claim that the “small-k opportunity” is “invalidated” overstate what the results show. I would suggest making more rigorous claims.

  1. The empirical evidence for the key claims is weak.

The experiments use only two real-world datasets, with at most three protected groups and six scoring dimensions. While these settings demonstrate computational improvements, they provide limited evidence about how the methods behave under more challenging group overlap, large tie sets, or increasing numbers of protected groups. These factors are particularly important because the theoretical analysis identifies group-membership complexity as a central computational bottleneck.

The experiment in Section 6.2 is the sole empirical support for the motivation of utility loss, and it compares stabilized utility-loss solutions against unstabilized w-difference solutions. Any minimizer of a distance to an infeasible w0w^0 over a closed feasible region will lie on the boundary. A margin-constrained w-difference variant is an obvious fair comparison, but it is missing.

Questions for authors
  1. How would the proposed framework ensure that a deployed selection system actually selects the fair subset? Does the algorithm return an explicit tie-breaking policy or selected subset along with the scoring weights? How would the guarantees change under a fixed or randomized tie-breaking rule?
  2. The theoretical efficiency of the verification algorithm relies on a constant number of protected groups, with the backtracking complexity depending on the number of distinct membership profiles. Can you provide additional runtime results varying the number of protected groups, the number of nonempty membership profiles, and the number of candidates tied at the selection boundary? This would help determine how broadly the proposed efficiency advantages extend beyond the relatively small-group settings evaluated.
  3. Since ATC+ ignores ties, how many of its outputs in the multi-dimensional experiments are actually unfair or suboptimal once ties are accounted for? Can you clarify how this affects the validity and interpretation of the reported runtime comparisons?

Suggestion (not affecting my score):

I would suggest keeping a related work section in the main content. You can make it short if space is limited and present a longer version in the appendix.

4 comments

We thank the reviewer for their comments and questions. Please find our answers below.

W1. Unclear practical usage and limited novelty relative to Cai.

Unclear practical usage: We agree in many real-world selection systems, tie-breaking may be determined by a fixed or randomized procedure. However, such a tie-breaking rule may not necessarily be designed with fairness considerations in mind, and previous work (Cachel & Rundensteiner FAccT'25; Cai SoCG'26) has already argued that such a tie-breaking rule may lead to reduced fairness. In fact, since our algorithm for the Fair Top-k Verification problem explicitly searches for a fair tie-breaking outcome, it can be readily adapted to return a fair top-k subset given a fair scoring function. Such an adaptation has already been implemented since our post-processing approach to stabilize a fair weight vector requires a target fair top-k subset as input.

Limited novelty relative to Cai: We agree our algorithmic contributions may appear to be incremental augmentations of the prior work, and that the NP-hardness result may appear "unsurprising", if one "reads our work by extracting the locally defined contribution of each individual component" (line 483 of our submission). However, as we have argued throughout our submission, the primary significance of these results lies in their interconnection, rather than their locally defined contribution. Specifically, we formally establish the connection between the Fair Top-k Selection problem and the combinatorial covering problem (in particular, Set Cover) to prove its NP-hardness. This result delivers a critical message: the number of protected groups can affect runtime efficiency, thereby correcting an implication of previous work. Moreover, our conditional fine-grained lower bound for small k is not established through a combinatorial covering problem. More importantly, this lower bound provides a key structural insight that guides our efficient algorithm for the Fair Top-k Verification problem when both the number of protected groups and k are small (see Remark 2 in our submission). The algorithm is a key component to augment the k-level based algorithm to recover its (theoretical) efficiency for small k, while also constituting an algorithmic contribution in its own right through its important role in experimental exploration. Furthermore, our augmentations yield new structural insights such as a monotonicity property of the utility of the top-k subset (Section 5.2.1), and many of our augmentations consider practical engineering considerations, with some of the design choices informed by our experimental observations (a few examples are highlighted in Appendix A). Thus, we believe the true crux of this work lies in how these components are interconnected. Evaluating these components in isolation would therefore miss the main contribution of our work. "Ultimately, the whole is greater than the sum of its parts" (line 485 of our submission).

Since establishing NP-hardness under the worst-case scenario is standard practice, we do not explicitly state it in our "even in 2-D" framing. We will clarify this in the revised version. As for the claim that the "Small k Opportunity" is "invalidated", we do state it as a worst-case result on lines 77–78: "invalidating the 'Small k Opportunity' identified in (Cai, 2026a) in the worst case". Nevertheless, we will make this claim more clear in the revised version.

Q1: How would the proposed framework ensure that a deployed selection system actually selects the fair subset? Does the algorithm return an explicit tie-breaking policy or selected subset along with the scoring weights? How would the guarantees change under a fixed or randomized tie-breaking rule?

As in our reply to W1 (Unclear practical usage), our algorithm for the Fair Top-k Verification problem can be easily adapted to return a fair top-k subset, and it has been implemented in our work. So, the fair subset can be easily obtained from the returned scoring weights with one additional step. Of course, one thing to note is that it is possible to have multiple tie-breaking outcomes satisfying the fairness constraints and the utility loss objective can help narrow down the choices (see Example 3 in Appendix D). However, multiple plausible tie-breaking outcomes may still remain. We believe users should make the final choice based on their own contextual understanding of fairness. In contrast, a fixed or randomized tie-breaking rule will not guarantee to produce a top-k subset satisfying the fairness constraints, thus resulting in decreased fairness.

W2. The empirical evidence for the key claims is weak.

Our empirical evaluation focus on real-world datasets, where the number of protected groups is indeed small. Also, as the reviewer may have noticed, our hardness results are established considering the worst-case scenario. So, the group-membership complexity may be a central computational bottleneck in theory without necessarily being one in practice. Indeed, one of our contributions is an efficient algorithm for more practical settings where both the number of protected groups and k are small, guided by a structural insight from our hardness analysis. As an additional note, we do not use synthetic datasets in our experiments as explained in Appendix H.1 (lines 1446–1449), following (Cai SoCG'26).

Also, in the Appendix F.3, line 1131–1133, we have already discussed the reason why w-difference solutions can not be stabilized in the same way as the utility-loss solutions: "the process of obtaining a stable fair weight vector cannot be applied to a fair weight vector with the minimum w difference since the resulting stable fair weight vector may have a larger w difference (see Figure 2 for an illustration). In contrast, for a fair weight vector with the minimum utility loss, the stabilization does not change the utility loss". Thus, we believe the margin-constrained w-difference variant is not a valid alternative, as it will fail to produce a fair weight vector that minimizes the w difference. We will move this discussion to the main text in the revised version. Moreover, we would like to mention that stability is not the sole motivation for the utility loss objective, as the utility-fairness trade-off is an important topic in algorithmic fairness.

Q2. The theoretical efficiency of the verification algorithm relies on a constant number of protected groups, with the backtracking complexity depending on the number of distinct membership profiles. Can you provide additional runtime results varying the number of protected groups, the number of nonempty membership profiles, and the number of candidates tied at the selection boundary? This would help determine how broadly the proposed efficiency advantages extend beyond the relatively small-group settings evaluated.

In the Appendix F.1.1, we have a detailed runtime analysis of the verification algorithm, which is further refined in the Appendix G.1 regarding our practical improvements. In short, the algorithm runs in O(n⋅d+tβ−1)O(n\cdot d + t^{\beta - 1}) time, where t≤kt\leq k is the number of positions to be filled from the set of candidates tied at the top-k selection boundary and β≤2np\beta \leq 2^{n_p} is the number of nonempty membership profiles, with npn_p denoting the number of protected groups. Notably, the tβ−1t^{\beta-1} term is absorbed when k=polylog(n)k=\text{polylog}(n) and β\beta is a constant. The cost associated to the number of candidates tied at the selection boundary is negligible as our algorithm only perform a linear scan to convert the set of such candidates into an array storing the number of candidates for each nonempty membership profiles, whose cost is absorbed by the O(n⋅d)O(n\cdot d) term. Our empirical evaluation focuses on the small-group setting for two reasons: we aim to assess practical performance on real-world datasets, which typically have few protected groups, and our verification algorithm is specifically designed for such a scenario.

Q3. Since ATC+ ignores ties, how many of its outputs in the multi-dimensional experiments are actually unfair or suboptimal once ties are accounted for? Can you clarify how this affects the validity and interpretation of the reported runtime comparisons?

In Appendix H.1 (line 1457), it has been stated that "the tie-breaking was ignored in the implementation of ATC+, as it was not obvious how to do so without introducing significant runtime overhead for ATC+". Making ATC+ correctly handles ties requires additional computational steps, increasing its run time. This means our performances advantages will be even greater when compared with the corrected ATC+ variant. Given that our experiments have already demonstrated speedups of several orders of magnitude, we believe these results sufficiently support our runtime efficiency claim over ATC+. Furthermore, given ATC+'s clear runtime inefficiency relative to our method, we use it solely to demonstrate the runtime advantage and do not evaluate its solution quality.

Suggestion 1 (not affected score). I would suggest keeping a related work section in the main content. You can make it short if space is limited and present a longer version in the appendix.

We will certainly do so in the camera-ready version, should that opportunity arise.

Reference

Kathleen Cachel and Elke Rundensteiner. Group fair rated preference aggregation: Ties are (mostly) all you need. FaccT 2025.

Guangya Cai. Finding a fair scoring function for top-k selection: From hardness to practice. SoCG 2026.

Thank you for the response and revisions. The clarification that the implementation can recover an explicit fair top-kk subset addresses my deployment question. Also I appreciate the clearer worst-case framing. However, I am still not fully convinced by the limited empirical evidence. And I still think the algorithms are largely incremental extensions of Cai’s framework. I will remain my rating.

We thank the reviewer's eventual reply.

As we have stated in our previous response, we agree the algorithms are largely incremental extensions if one "reads our work by extracting the locally defined contribution of each individual component and ignoring its nonlinear structure (Figure 1)" (Conclusion of our submission). However, as we have argued throughout the submission (Introduction, Figure 1, Conclusion, and Appendix A, etc.), the main contribution of the submission is not the locally defined contribution of individual components, but rather the integrative whole. Given our algorithmic contributions are tightly interconnected with our hardness analysis, practical engineering, and empirical observations, we respectfully disagree with the reductive evaluation of our work.

There is a reason we chose "An Integrative Approach" as the title of our work, as it captures the central argument developed throughout the submission. We sincerely hope that readers will recognize the integrative approach for what it is: an integration that goes clearly beyond a mere augmentation of Cai's two-pronged solution.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Revised submission

We thank all the reviewers once again for their feedback. We have uploaded a revised version of the submission, with all changes highlighted in blue. Below, we summarize the revisions made:

  1. Clarification of the hardness results: We have further clarified that the hardness results are established in the worst-case setting (Introduction, lines 76 and 78–79), addressing Reviewer 44QL's comment.
  2. Relocation of a discussion on the stabilization method: We have moved the discussion explaining why the stabilization method cannot be applied to a fair weight vector that minimizes the w difference from Appendix F.3 to Section 4.2 (lines 288–292), addressing Reviewer 44QL's comment.
  3. Condition for a strictly positive margin: We have added a discussion of the condition under which the margin ξ in Equation (5) is strictly positive (Section 4.2, lines 282–288), addressing Reviewer apby's comment.
  4. Clarification of the stability experiments: We have further clarified the objective of the stability experiments and the procedure used to generate perturbed weight vectors (Section 6.2, lines 445–450), addressing Reviewer KNjT's comment.

As suggested by Reviewer 44QL, we also plan to incorporate a related work section into the main content in the camera-ready version, should the opportunity arise.

Finally, please note that the line numbers cited in our previous response may differ from those in the revised submission, as they referred to the original version. Nevertheless, all the necessary context and explanations have been incorporated into our previous response.

As the discussion period is approaching its end (Nov 18), we look forward to hearing back from the reviewers soon.

AuthorsOpenReview ↗

Final remark

As the discussion period is soon coming to an end, we would like to offer a few final words that we believe may be beneficial to future readers who may come across this discussion.

Anyone who has read this version of the paper in its entirety may have noticed some unusual statements in Introduction and Conclusion. We are fully aware that these statements will not prevent some readers from reading our work through a siloed, reductive, and component-wise lens. Nevertheless, as author(s), we remain hopeful that readers will find in this work the same intellectual interest that we do. In our view, "novelty" should not necessarily be defined in such a narrow way.

As stated in one of our previous responses, there is a reason we chose An Integrative Approach as the title of this work. We sincerely hope readers will reflect on what this title truly entails while reading the paper.

Once again, as we have argued repeatedly throughout our paper and this discussion, but still feel the need to say it one last time: Ultimately, the whole is greater than the sum of its parts.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.