acceptodds
Under review as a conference paper at ICLR 2027

Beyond a Single Sparse Solution: Certifying Feature Necessity under Inexact Multi-Label Supervision

Abstract

Feature selection under inexact multi-label supervision is complicated by both ambiguous label assignments and redundant predictive supports. Existing methods typically identify the support of a single fitted solution, which may be predictive but does not establish whether its selected features are necessary. We study feature necessity over the family of near-optimal predictors. Our framework evaluates a feature set through counterfactual deletion: the set is removed, all remaining variables are re-optimized, and the resulting deletion cost determines whether any comparably good predictor can avoid it. This formulation characterizes both individual necessity, where a feature must appear in every near-optimal predictor, and collective non-removability, where individually replaceable features cannot be removed jointly. We instantiate the framework with a convex ambiguity-aware objective that preserves set-valued supervision and induces row-sparse feature selection. Strong convexity yields safe bounds on optimization error and deletion costs, enabling efficient screening and selective re-optimization, while monotonicity supports structured search for collective relevance. Extensive experiments on controlled and real-world datasets validate the predicted individual and collective structures and demonstrate the practical effectiveness of the resulting certificates and feature rankings.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Inexact Multi-Label Supervision, Feature Selection, Feature Necessity, Near-Optimal Predictors, Counterfactual Deletion

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_5pwnOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FIND, Feature necessity under INexact supervision via Deletion—to study feature necessity over the family of near-optimal predictors. Feature selection is an important topic, and the idea proposed in this paper is interesting. The authors made a series of theoretical research and some empirical experiment results presented.

Critical strengths

--- important topic, and idea is interesting. --- good theoretical induction and analysis

Critical weaknesses

The paper does not make enough empirical experiments. Compared with the theoretical side, the experiments and results are weak.

The current results are insufficient to demonstrate the effectiveness of the proposed method. It would be better if you can provide more empirical evidence to demonstrate the usefulness of the proposed method.

Questions for authors

Could you provide more direct and explicit experiments and results to show clearly the evidence of your proposed method is better?

Could you please use two datasets as examples to clearly show which feature sets are really critical? what consequences were generated by removing them created ?

3 comments

1. Predictive performance across benchmark settings

We ranked FIND and the eight baselines within each of the 18 benchmark settings using the budget-averaged scores reported in the paper. Ties at the reported precision received average ranks. FIND achieved the best mean rank on all four metrics.

MethodRL rankCE rankAP rankMicro-F1 rank
FIND1.0561.0001.1111.000
PML-FSLA2.2222.6113.0565.861
PML-FSMIR5.2784.9724.7225.667
PML-FSSO3.5563.4724.0565.750
fPML5.3895.2225.8894.722
PML-LD4.8335.1394.0005.250
PML-FP8.5567.6678.1397.417
PML-MAP7.3337.8897.8893.667
PAMB6.7787.0286.1395.667
MetricFIND ranked firstFriedman statistic (df = 8)Raw pHolm-adjusted p
Ranking Loss17/18112.4741.18e−204.72e−20
Coverage Error18/18104.9274.19e−191.26e−18
Average Precision16/1898.8277.41e−181.48e−17
Micro-F118/1864.9894.85e−114.85e−11

The Friedman tests assess overall differences among the nine methods. Because some settings share an underlying dataset, we treat the setting-level p-values as exploratory. The average ranks and first-place counts describe the consistency of FIND's predictive performance across the evaluated settings.

2. Controlled ablation of necessity-first ranking

We held the fitted selector, splits, budgets, and downstream classifier fixed and compared FIND's necessity-first ranking with pure row-norm ranking (Norm). Prioritizing certified necessary features changed the selected top-kk set in 323 of 1,440 fold–budget pairs (22.4%). The table includes all three settings with changes in at least 75% of pairs. Scores average all 80 pairs per setting.

SettingChanged pairsAP ↑ NormAP ↑ FINDMicro-F1 ↑ NormMicro-F1 ↑ FIND
Music emotion77/80 (96.3%)0.5740.5800.4840.488
Music style61/80 (76.3%)0.7080.7090.2660.267
Enron (c=9c=9)60/80 (75.0%)0.4170.4240.2750.282

In the three settings where necessity-first and pure row-norm ranking differed most, our method demonstrates advantages in both AP and Micro-F1. These results show that prioritizing certified necessary features can improve downstream prediction when it substantially changes the selected feature set, with gains varying by setting and metric. In the seven settings where both rules selected identical sets at every evaluated budget, row-norm ranking from the fitted ambiguity-aware objective reproduced FIND's predictive performance. The same objective accommodates candidate-label ambiguity and defines the near-optimal family used for deletion certification. In these seven settings, the predictive results are therefore attributable to the shared base selector.

3. Two examples of certified features and deletion outcomes

We illustrate individual necessity on Music emotion and collective non-removability on Mediamill (c=3c=3), using evaluation fold 2 in both cases. Feature indices are one-based. Starting from FIND's top-kk subset, we remove the certified features, refill the subset from the remaining ranked features to keep kk fixed, and retrain the same downstream classifier.

Dataset and budgetCertified featuresMetricOriginalAfter deletion
Music emotion, 10% (k=10k=10){1, 27, 59, 66, 68}RL ↓0.2700.278
Micro-F1 ↑0.4910.482
Mediamill (c=3c=3), 5% (k=6k=6){38, 41, 53}RL ↓0.1120.116
Micro-F1 ↑0.5880.567

On Music emotion, FIND certified each of the five features as individually necessary. Replacing the entire core increased held-out Ranking Loss and reduced Micro-F1 by 0.91 percentage points at the displayed budget.

On Mediamill, each of the three features was certified individually removable. Their joint deletion had a certified objective increase of at least 4.6815×10−44.6815\times10^{-4}, above the tolerance upper bound of 3.1307×10−43.1307\times10^{-4}. Thus, every near-optimal predictor must retain at least one member of this group. At the displayed budget, replacing the group increased held-out Ranking Loss and reduced Micro-F1 by 1.92 percentage points. The certificate establishes collective non-removability under FIND's learning objective, while the test scores show the predictive consequence at this budget.

Many thanks for your providing responses to my comments. While I can see your great effort, you did not clearly mention the details of the four comparison methods. I would like to see who developed those comparison methods, in which year, which journals and/or conferences published them, and the authors.

I would also like to see how many features and what features those methods automatically selected, and the running time of the those methods.

Without those details, I cannot judge whether FIND can actually perform better than the most recent or state-of-the-art methods. So for now, I keep my scores unchanged. I can reconsider this if you provide those details.

Thank you for your reply. All details of the comparison methods are included in Appendix B, Experimental Details. Where can you get the details you want.

Reviewer_e35gOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies feature selection under inexact multi-label supervision—covering partial multi-label learning (PML, where each instance is annotated with a candidate label superset si ∈ {0,1}q) and complementary multi-label learning (CML, via si = 1 - ci). The authors argue that existing PML feature selectors jointly fit pseudo-labels and a sparse projection matrix, yielding only a single admissible support rather than establishing whether selected features are indispensable across all near-optimal explanations.

To address this, the paper proposes FIND (Feature necessity under INexact supervision via Deletion):

Convex Ambiguity-Aware Objective: Instead of estimating pseudo-labels, FIND retains the convex hull of admissible binary label vectors Qi = conv(Yi) = {z ∈ [0, 1]q : 0 ≤ z ≤ si, 1T z ≥ 1} and minimizes the Moreau-envelope squared distance ½dist2(WT xi + b, Qi) with row-sparse ℓ2,1 regularization λ ‖W‖2,1 and strongly convex ridge regularization γ/2(‖W‖F2 + ‖b‖22) (Eq. 4). Counterfactual Deletion & Rashomon-Set Necessity: Given the unconstrained optimum F^⋆ and intercept-only optimum F0^⋆, FIND defines the near-optimal (Rashomon) family 𝒜ϵ = {(W, b) : F(W, b) ≤ F^⋆ + ϵ} with relative tolerance ϵ = η (F0^⋆ - F^⋆). For any feature subset G, the counterfactual deletion cost D(G) = minW_(G,:)=0, b F(W,b) - F^⋆ satisfies D(G) > ϵ ⇔ ‖WG,:‖F > 0 for all (W, b) ∈ 𝒜ϵ (Theorem 1). Singleton tests G={k} identify the individual necessity core Nϵ = ⋂(W,b)∈𝒜_ϵ supp(W), while sets G where D({k}) ≤ ϵ (∀ k ∈ G) but D(G) > ϵ characterize collective non-removability. Safe Screening & Prefix Search: Using γ-strong convexity and the minimum-norm subgradient residual r = dist(0, ∂F(Ŵ, b̂)), Theorem 2 derives certified intervals [D̲(G), D̄(G)] and [ϵL, ϵU] to screen singletons and trigger constrained proximal-gradient re-optimization only when unresolved. When the certified individual core N̂ is empty, exponential bracketing and binary search over nested prefixes of R̂ (ordered by decreasing ‖Ŵk,:‖2) locate the first collectively non-removable prefix.

Experiments on controlled synthetic data (Table 1) and 18 PML settings across 5 controlled multi-label datasets (Birds, Corel5k, Enron, HumanPseAAC, Mediamill) and 3 real-world PML datasets (Music emotion, Music style, Mirflickr) show that FIND recovers synthetic individual/collective structures and achieves the best downstream linear-SVM performance in 69 of 72 dataset–metric comparisons across Ranking Loss, Coverage Error, Average Precision, and Micro-F1.

Critical strengths

Principled Convex Formulation of Set-Valued Supervision and Exact Rashomon-Set Necessity Certificates: Formulating partial/complementary multi-label learning via the squared Euclidean distance to the admissible label polytope Qi = conv(Yi) (Proposition 3) is clean and well-motivated. Because Euclidean projection onto Qi reduces to box clipping followed (when 1T clip(pCi, 0, 1) < 1) by projection onto the unit simplex Δi, the resulting Moreau envelope has a 1-Lipschitz continuous gradient (Proposition 4). This completely avoids the non-convex alternating label-disambiguation heuristics used in prior PML selectors and makes rigorous Rashomon-set certification possible. Combining counterfactual re-optimization D(G) = F-G^⋆ - F^⋆ with computable subgradient-residual duality/strong-convexity bounds (g = r2 / (2γ), Theorem 2) provides mathematically sound, solver-tolerance-aware certificates without assuming exact optimization convergence. Comprehensive Predictive Evaluation Across 18 PML Settings and 4 Metrics: The empirical evaluation in Section 4 and Appendices B–C is thorough and clearly reported across 18 benchmark settings, 8 PML baselines, 20 feature budgets (1%–20%), and 4 multi-label metrics with full budget trajectories (Figures 3–6), achieving top performance in 69/72 dataset–metric aggregations.

Critical weaknesses

Unablated Attribution of Predictive Gains: Base Convex Objective (‖Ŵk,:‖2) vs. Counterfactual Deletion Certification: While the paper's central thesis is moving beyond a single sparse solution, the predictive ranking π evaluated in Tables 2 and 6–8 is constructed by placing certified necessary features N̂ first and ordering both N̂ and the remaining features R̂ by the row norms ‖Ŵk,:‖2 of the single fitted base solution Ŵ (Section 3.3, Algorithm 1 Line 27). Crucially, in the 4 benchmark settings where N̂ = ∅ (e.g., Mediamill c=3,4,5 and HumanPseAAC c=5), FIND's feature ranking π is 100% identical to simply ranking features by ‖Ŵk,:‖2 from the single sparse solution of Eq. (4). Even in the 14 settings where N̂ ≠ ∅, features with D({k}) > ϵ are typically those that already have the largest row norms ‖Ŵk,:‖2 in Ŵ. Because the paper does not include an ablation comparing: Base Eq. (4) only (ranking purely by ‖Ŵk,:‖2 of the single fitted solution Ŵ without deletion certification), Full deletion-cost ranking (ranking all features directly by their re-optimized singleton deletion cost D({k})), and FIND's two-block hybrid ranking (N̂ then R̂, each sorted by ‖Ŵk,:‖2), it is impossible to determine whether the large predictive improvements over the 8 baselines in Table 2 stem from the counterfactual necessity certification or simply from the new convex loss F(W,b) in Eq. (4). Restrictive Collective Discovery Design, Loose Lower-Bound Screening, and Evaluation Protocol Questions: Mutual Exclusivity of Individual and Collective Discovery: In Algorithm 1 (Line 28), collective group discovery is executed only if N̂ = ∅ and Û = ∅. If even a single feature in the dataset is individually necessary (|N̂| ≥ 1), FIND aborts group search altogether. In real-world problems, the presence of one indispensable feature does not preclude other features from forming redundant correlated groups (collinear feature pairs/clusters) within R̂. Prefix-Only Collective Sets and Matched-Deletion Confounding: When N̂ = ∅, Algorithm 1 only tests nested prefixes Gj = {r1, …, rj} ordered by ‖Ŵrj,:‖2. If the redundant signal copies lie at ranks r5 and r6, the returned certificate G6 = {r1, …, r6} includes four unrelated features rather than isolating a minimal non-removable group. Furthermore, in the matched-deletion experiment (Table 3) for collective sets, deleting the prefix Gj_U (which is literally the top-jU highest-norm features in Ŵ) from the top-k subset and comparing it against deleting jU random features from top-k only shows that the top-jU features are more predictive than lower-ranked features in top-k, rather than validating synergistic group substitutability. Tightness of the Initial Lower Bound D̲(G) and Runtime Overhead: The safe lower bound D̲(G) = γ/2(‖ŴG,:‖F - R)+2 (Eq. 20) scales with the small ridge parameter γ (γ/L ∈ {10-4, 10-3, 10-2}), ignoring both the data loss ℓi and the ℓ2,1 penalty λ ‖W‖2,1. The paper does not report what fraction of features are actually certified as necessary by D̲({k}) > ϵU without re-optimization versus how many require constrained proximal-gradient solves, nor does it report wall-clock runtimes relative to baselines. Scale and Modern Representation Scope: All experiments are restricted to small tabular/classical benchmarks (d ≤ 499, q ≤ 20) with linear models; demonstrating FIND on higher-dimensional settings or linear/sparse probing over pretrained foundation model embeddings under noisy/ambiguous multi-label supervision would substantially broaden its practical impact.

Questions for authors

Clarification / Core Assessment — Ablation isolating Base Eq. (4) (‖Ŵk,:‖2) vs. Deletion Certification: In Table 2 (and Tables 6–8), how much does the predictive ranking actually change when placing N̂ before R̂ compared to simply ranking all features by ‖Ŵk,:‖2 of the single fitted solution Ŵ of Eq. (4)? What is the performance of (a) pure ‖Ŵk,:‖2 ranking from Eq. (4) and (b) ranking all features by exact singleton deletion cost D({k})? Clarification / Core Assessment — Screening Rate and Computational Cost: Across the 18 benchmark settings at η = 0.01, what percentage of singletons in N̂ and R̂ are certified directly by the initial bounds [D̲({k}), D̄({k})] in Eq. (20)–(21), and what percentage require running the constrained solver in Line 12 of Algorithm 1? What is the total wall-clock runtime of FIND compared to the 8 baselines? Clarification / Core Assessment — Use of Clean Validation Targets (Appendix B, Line 723): Line 723 states that one of the five partitions "is reserved for hyperparameter selection using clean validation targets". Since clean ground-truth labels are unavailable in true PML settings, were clean validation targets used to tune hyperparameters for both FIND and all eight baselines, and how does FIND perform when (λ, γ) are selected using only the inexact candidate-label validation loss? Suggestion for Future Work — Post-Pruning of Collective Prefixes and Coexisting Group Search: Even when N̂ ≠ ∅, conditioning on N̂ (or searching within R̂) and applying a backward-deletion pass on the certified prefix Gj_U could isolate minimal collectively non-removable subsets rather than returning full prefixes.

1 comment

1. Predictive ranking and the role of certification

We held the fitted selector fixed and compared FIND's necessity-first ranking with pure row-norm ranking (Norm), using identical folds, feature budgets, and downstream classifiers. Across 18 settings, prioritizing the individual core changed the selected top-kk set in 323 of 1,440 fold–budget pairs (22.4%). The table reports all three settings where the selected set changed in at least 75% of pairs. Each metric is averaged over all 80 pairs in that setting.

SettingChanged pairsAP ↑ NormAP ↑ FINDMicro-F1 ↑ NormMicro-F1 ↑ FIND
Music emotion77/80 (96.3%)0.5740.5800.4840.488
Music style61/80 (76.3%)0.7080.7090.2660.267
Enron (c=9c=9)60/80 (75.0%)0.4170.4240.2750.282

In the three settings where necessity-first and pure row-norm ranking differed most, our method demonstrates advantages in both AP and Micro-F1. These results show that prioritizing certified necessary features can improve downstream prediction when it substantially changes the selected feature set, with gains varying by setting and metric. In the seven settings where both rules selected identical sets at every evaluated budget, row-norm ranking from the fitted ambiguity-aware objective reproduced FIND's predictive performance. The same objective accommodates candidate-label ambiguity and defines the near-optimal family used for deletion certification. In these seven settings, the predictive results are therefore attributable to the shared base selector.

Direct ranking by singleton deletion cost would test whether that cost is a useful predictive score. The paired ablation above isolates the contribution of core prioritization to the ranking used in the paper. Our certificates determine whether a deletion cost crosses the tolerance. Costs on the same side of the tolerance can have overlapping intervals, so certification can finish before their relative order is determined. We have not evaluated deletion-cost ranking, which would require further refinement where such uncertainty affects feature selection.

2. Screening and computational cost

At η=0.01\eta=0.01, the initial bounds resolved 13,752 of 16,352 singleton checks (84.1%) across 72 evaluation folds. The remaining 2,600 checks required constrained re-optimization, averaging 36.1 calls per fold. Three checks remained unresolved.

MethodMean recorded time per fold (s)Timing scope
FIND19.13Archived fold duration
PML-FSLA19.06Archived fold duration
PML-FSMIR12.38Archived fold duration
PML-FSSO8.32Archived fold duration
fPML22.57Archived fold duration
PML-FP9.26Archived fold duration
PML-MAP27.59Archived fold duration

For each baseline with timing records, we averaged the available folds within each setting and weighted the 18 settings equally. We report the archived values as recorded.

3. Hyperparameter selection using candidate labels In the main comparison, clean validation labels were available for hyperparameter selection to FIND and all eight baselines. To assess FIND when clean validation labels are unavailable, we compared clean-label validation with candidate-label validation using the same nine-point grid for both variants.

Validation criterionRL ↓CE ↓AP ↑Micro-F1 ↑
Clean labels0.1760.2610.6010.420
Candidate labels0.1790.2640.6010.419

Candidate-label tuning was slightly worse on average across all four metrics, with larger drops in some settings. Both variants used the original split protocol, in which tuning-stage fits had access to features and candidate labels from later evaluation folds.

4. Collective discovery and matched deletion

The current search targets nested ranked prefixes when the singleton tests certify that every feature is individually removable. Individually necessary features can also coexist with substitutable groups. Searching for such groups and applying backward pruning to certified prefixes are natural extensions.

The matched-deletion experiment compares held-out performance after removing the same number of features at a fixed feature budget. The individual and joint deletion certificates establish the collective property: each member can be omitted individually, but no near-optimal predictor can omit all members simultaneously.

Evaluation scope. Our empirical results cover linear predictors on the reported benchmarks. Evaluating FIND on higher-dimensional pretrained embeddings is an open direction.

Reviewer_BedPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on an important issue in deep learning, the feature necessity when multi-label supervision is not guaranteed to be completely accurate, since precise annotation is expensive in real-world cases. This paper analyses the feature necessity under set-valued multi-label supervision with a model with superior performance, and thus determines whether a selected feature is necessary. Then, the authors further develop a set-deletion framework to analyze individual and collection importance. Finally, extensive experiments on real-world datasets validate the effectiveness of the proposed method.

Critical strengths
  1. This paper focuses on an interesting problem of analyzing feature necessity with inaccurate multi-label supervision. Specifically, several techniques such as the near-optimal model family and the set-deletion framework with several mathematical tools are adopted. The paper is supported by several mathematical equations and theoretical analysis, with the detailed algorithm and proof provided in the appendix.
  2. The model shows significant improvement in performance compared with existing methods. The paper additionally conducts extensive experiments on the sensitivity of the model to different parameters.
  3. Extensive theoretical proof of the proposed method is provided to support the effectiveness and design of the overall framework.
Critical weaknesses
  1. The main contribution of this paper is not well-stated in the introduction, making it difficult to follow how the proposed method could address this important problem effectively. Additionally, the paper seems to be a combination of existing techniques, and the main difference from other works and the major contribution could be better revised and analyzed in detail in the introduction to support the novelty of this paper.
  2. Figure 1 seems to be informal, the equations are pure text instead of mathematical equations, and the hat symbol in part 03 is not in the correct position. Moreover, it is also difficult to find the correspondence between each section and the modules in the figure, since they are named differently in the Figure and in the sections. The author should proofread and revise the paper following the three stages shown in the Figure.
  3. The paper contains extensive equations in the method section, while lacking detailed explanations of them, making it hard to follow or verify their correctness. Additionally, some key ablation studies are not provided. For example, what is the effect of removing or replacing the ambiguity-aware feature selector and the individual core on the model performance? Moreover, could the authors also label the second-best results in Figure 2 to provide a clearer illustration of the model improvements compared with existing methods?
Questions for authors

Please refer to weak points.

1 comment

1. Main contribution and distinction from prior work

We agree that the introduction should state the problem and contribution more explicitly. Existing partial multi-label feature selectors typically return the support of one fitted model. Under ambiguous labels and correlated features, several predictors can fit the observations comparably well while relying on different feature sets. For example, if two features are substitutable, either can be absent from a near-optimal predictor, yet deleting both can rule out all near-optimal predictors. A single fitted support cannot distinguish these possibilities.

FIND addresses this gap by asking whether any near-optimal predictor can avoid a given feature set. It compares the optimal objective before and after disabling that set and re-optimizing the remaining parameters. This comparison distinguishes individual necessity from collective non-removability under the same set-valued supervision criterion. The computational components draw on established convex optimization tools. Our contribution is the formulation of feature necessity under inexact supervision and the use of computable bounds to certify it despite finite solver accuracy. We will revise the introduction to make this distinction explicit and use the example above to connect the problem, the proposed test, and the two types of certificates.

2. Figure 1 and its correspondence with the method

We will replace the plain-text equations with typeset mathematics, correct the misplaced hat in Stage 03, and check the symbols against their definitions in the paper. We will also align the stage names with the method headings and add the corresponding section references to the caption. The revised figure will show how the admissible label sets define the learning objective, how that objective defines the near-optimal family and deletion test, and how the certified bounds yield individual or collective conclusions. We will use the same terminology in the text so that readers can follow the figure as a guide to the method.

3. Equation explanations, ablation, and second-best results

Equation explanations. We will add short explanations at three points in the method. First, the admissible target set lets the loss compare each prediction with the closest target consistent with the observed annotation. Second, re-optimizing after deletion lets the remaining features compensate for the removed ones, so the deletion cost measures the best attainable fit under the constraint. Third, we will explain how the bounds determine the decision. A lower bound above the tolerance's upper bound certifies non-removability. An upper bound at or below the tolerance's lower bound certifies removability. If neither condition holds, the initial bounds leave the decision unresolved and constrained re-optimization can refine it.

Individual-core ablation. We held the fitted selector fixed and compared FIND's necessity-first ranking with pure row-norm ranking (Norm), using identical folds, feature budgets, and downstream classifiers. Across 18 settings, prioritizing the individual core changed the selected top-kk set in 323 of 1,440 fold–budget pairs (22.4%). The table reports all three settings where the selected set changed in at least 75% of pairs. Each metric is averaged over all 80 pairs in that setting.

SettingChanged pairsAP ↑ NormAP ↑ FINDMicro-F1 ↑ NormMicro-F1 ↑ FIND
Music emotion77/80 (96.3%)0.5740.5800.4840.488
Music style61/80 (76.3%)0.7080.7090.2660.267
Enron (c=9c=9)60/80 (75.0%)0.4170.4240.2750.282

In the three settings where necessity-first and pure row-norm ranking differed most, our method demonstrates advantages in both AP and Micro-F1. These results show that prioritizing certified necessary features can improve downstream prediction when it substantially changes the selected feature set, with gains varying by setting and metric. In the seven settings where both rules selected identical sets at every evaluated budget, row-norm ranking from the fitted ambiguity-aware objective reproduced FIND's predictive performance.

Ambiguity-aware selector. Both ranking variants use the same ambiguity-aware objective. The ablation shows how well the base selector performs with row-norm ranking and isolates the incremental effect of core prioritization. Isolating the effect of retaining the admissible target set would require replacing it with a fixed target while keeping the rest of FIND unchanged.

Second-best results. We interpret the request as referring to the predictive comparison in Table 2. We will mark the second-best result for each metric and setting in Table 2 and all other predictive comparison tables, including those in the appendix.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.