PaperlayerView exact revision on arXiv

Source-linked AI summary

SelectArena: A Modern Benchmark for Feature Selection on Real-world Tabular Data

Bastian Schäfer, Dominika Matus, Lennart Purucker, Ricardo Knauer, Erik Rodner, Frank Hutter, Anton Sokolov, Arnaud Liefooghe

arXiv:2610.13644v1cs.LGcs.AI

TL;DR

Feature-selection benchmarks provide limited guidance despite practitioners’ need to balance performance, computation, and interpretability. SelectArena evaluates 15 methods on 71 curated real-world datasets with modern tabular models, finding metric trade-offs and identifying F-Test as a balanced choice.

  • Problem

    Existing feature-selection benchmarks provide limited practitioner guidance and often use limited datasets, unrealistic protocols, or outdated models despite the need to balance performance, computational cost, and interpretability.

    Source: §1 · p. 2

  • Method

    SelectArena evaluates 15 feature-selection methods on 71 manually curated real-world tabular datasets using standardized protocols and five downstream models, including a tabular foundation model.

    Source: §1 · p. 2

  • Results

    Tree-based RFImportance and CART and regularization-based ElasticNet and Lasso achieve the highest predictive performance; F-Test is the most balanced across four criteria, while search-based methods are slow and rank near the bottom on performance and stability.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Practitioners might choose RFImportance or CART when predictive performance is primary, and F-Test for interpretability-focused applications.

    Source: §5 · p. 9

    A one-hour time limit means slower methods may time out, and affected results reflect random fallback performance under that limit on the authors’ hardware.

    Source: §5 · p. 9

Abstract

from arXiv · show

Selecting a subset of features (columns) from a tabular dataset is a ubiquitous challenge in real-world applications. Industry experts seek to minimize model training costs, climate scientists require an interpretable subset of factors to understand extreme weather events, and clinicians need to choose among diagnostic tests due to financial and time constraints. When looking at the scientific literature to determine which feature selection method to use, practitioners across domains face an unrepresentative landscape of benchmarking. Despite extensive research, existing feature selection benchmarks provide limited guidance because they rely on outdated models, ad hoc evaluation protocols, or artificially generated data that do not reflect real-world problems. To close the gap between scientific benchmarking and real-world applications, we introduce SelectArena, a modern benchmark for feature selection. We compare 15 widely used feature selection methods on 71 manually curated real-world tabular datasets, covering classification and regression problems, using standardized evaluation protocols and state-of-the-art tabular (foundation) models. Our results reveal clear trade-offs across predictive performance, runtime, stability, and validity. Tree-based methods, CART and RFImportance, along with F-Test are Pareto-optimal when taking into account predictive performance and runtime. F-Test, information-theory-, and distance-based methods lead on stability and validity, with F-Test emerging as a balanced choice across our metrics. Search-based methods are slow, unstable and achieve low performance. SelectArena turns these findings into practical guidance and, as an extensible platform, supports future feature selection research.

1 INTRODUCTION

SelectArena addresses fragmented, outdated feature-selection benchmarking with a standardized comparison on curated real-world tabular data. Results show trade-offs across metrics, with F-Test emerging as a balanced option.

1 INTRODUCTION

  • SelectArena benchmarks 15 feature-selection methods on 71 manually curated real-world tabular datasets using state-of-the-art tabular models across predictive performance, runtime, stability, and validity.Source: §1 · p. 2
  • The benchmark dataset set was manually curated from an analysis of 368 datasets in feature-selection and machine-learning literature.The resulting datasets are released.Source: §1 · p. 2
  • The authors reviewed 296 feature-selection methods and selected 15 under a strict curation protocol.They provide their own implementations when well-maintained codebases are unavailable.Source: §1 · p. 2
  • A standardized evaluation pipeline assesses predictive performance, stability, validity, and runtime using five downstream models, including a state-of-the-art tabular foundation model.The pipeline code is open-source.Source: §1 · p. 2
  • Tree-based RFImportance and CART and regularization-based ElasticNet and Lasso achieve the highest predictive performance; F-Test is balanced across the four criteria, while search-based methods are slow and rank near the bottom on performance and stability.F-Test and GainRatio are fastest; F-Test, GainRatio, MI, and LaplacianScore lead on stability and validity.Source: §1 · p. 2

2 RELATED WORK

Prior feature-selection benchmarks are limited in methods, data, evaluation criteria, and use of modern downstream models. SelectArena contrasts with this landscape through broader coverage and a systematic benchmark comparison.

2 RELATED WORK

  • Benchmarks often examine no more than 10 methods or specific method subgroups, and commonly use synthetic, domain-specific, or non-tabular data; real-world tabular datasets usually number at most 20.Source: §2 · pp. 2–3
  • Most benchmarks focus on predictive performance, few assess stability or validity, and recent studies rarely include state-of-the-art downstream models such as tuned gradient-boosted trees or pretrained tabular foundation models.Source: §2 · p. 3
  • SelectArena benchmarks diverse methods across a broad dataset collection, evaluates performance, stability, and validity, and incorporates state-of-the-art models; Table 1 compares benchmarks by methods, datasets, models, and metrics.Source: §2 · p. 3, Table 1 · p. 3

3 SELECTARENA

SelectArena frames feature selection as choosing a useful subset under a feature-cost budget, then benchmarks curated methods and datasets using standardized evaluation protocols. The benchmark evaluates predictive performance, stability, validity, and runtime across multiple downstream models.

3.1 PROBLEM SETUP

  • The problem maximizes model-dependent subset utility subject to a total feature-cost budget, with uniform costs and budget k recovering the cardinality constraint |S| ≤ k.SelectArena mostly adopts cardinality constraints, which the paper describes as a simpler, more natural abstraction in many settings.Source: §3.1 · p. 4

3.2 BENCHMARKING SETUP

  • The benchmark curates 15 methods using tabular and task suitability, methodological independence, literature frequency, and subgroup coverage, and selects 71 real-world datasets from 368 candidates.Dataset criteria include at least 10 features, IID suitability for random splitting, and accessible source data.Source: §3.2 · p. 4
  • Five downstream models span diverse inductive biases, while the feature-selection routine is integrated into AutoGluon’s model-agnostic preprocessing pipeline with time limits and random fallback.The models are Linear, RF, LGBM, RealMLP, and TabICLv2; experiments comprised 79,875 performance runs.Source: §3.2 · pp. 4–5

3.3 EVALUATION

  • SelectArena evaluates methods on predictive performance, stability, validity, and runtime, varying feature budgets and applying a one-hour limit to selection and downstream-model experiments.For datasets with at most 100 features, k varies from ⌈p/6⌉ to ⌈5p/6⌉; larger datasets use an exponential regime, and random fallback handles methods that exceed constraints.Source: §3.3 · p. 6
  • Predictive performance uses ROC AUC, log-loss, or RMSE by task and is repeated across three seeds, while stability uses the Nogueira metric over 25 bootstrap samples.The Nogueira metric ranges from 0 to 1, with scores below 0.40 indicating poor stability and scores above 0.75 indicating excellent stability.Source: §3.3 · p. 6
  • Validity is assessed by adding uninformative Gaussian noise features at three proportions, then measuring selection precision as the share of selected features from the original set.Noise is added at 50%, 75%, and 100% of the original feature count; in this setting, selection precision equals normalized selection recall.Source: §3.3 · pp. 6–7

4 RESULTS

F-Test is the most balanced method across predictive performance, runtime, stability, and validity, although no method leads on every criterion. Results also show that method choice matters across predictors and that stability falls in low-data regimes.

4 RESULTS

  • RFImportance, CART, and F-Test lie on the predictive-performance/runtime Pareto front; search methods are slow and perform poorly, partly under the imposed time limit.These three methods also rank in the top four for performance and among the six fastest for runtime.Source: §4 · p. 7, Figure 4 · p. 7
  • Across model–method interactions, within-method improvement spread is at most 5.4pp, versus 15.8pp across methods under TabICLv2; method ordering is more consistent than model ordering.TabICLv2 ranks first for eight of 14 methods, while Linear ranks last for seven; Lasso and ElasticNet peak with Linear by only 0.5pp, whereas RFImportance and CART perform best with TabICLv2.Source: §4 · pp. 7–8
  • Most methods become less stable in low-data regimes; F-Test and distance-based methods rank highest, while F-Test and MI are the only high-performing methods also ranking high in stability.Search methods are only slightly more stable than random selection, which the authors attribute to high fallback rates.Source: §4 · p. 8, Figure 5 · p. 8
  • F-Test and information-theoretic methods achieve the highest validity, followed by distance-based methods; validity is less sensitive to samples per feature than stability.MarkovBlanket has lower validity than random, potentially because it favors removing redundant features common in real-world data.Source: §4 · pp. 8–9, Figure 6 · p. 8
  • F-Test comes closest to a balanced choice across all four axes, while RFImportance and CART have higher Elo ratings but lower stability.No method dominates every criterion; model-coupled search methods are competitive on none under the one-hour budget.Source: §4 · p. 9

5 CONCLUSIONS

SelectArena benchmarks 15 feature-selection methods on 71 manually curated real-world tabular datasets using five downstream models, including a tabular foundation model. The authors offer practical recommendations while noting limitations in the evaluation setting.

5 CONCLUSIONS

  • SelectArena evaluates 15 feature-selection methods on 71 manually curated real-world datasets spanning classification and regression, using five downstream models including a tabular foundation model.The authors describe it as the largest general feature-selection benchmark to date and the only one to include leading tabular models.Source: §5 · p. 9
  • Results are bounded by a one-hour runtime limit, fixed-size feature subsets, and stability and validity tests using bootstrapping and Gaussian noise.Timeouts can trigger a random fallback, and alternative evaluation metrics such as calibration or fairness may lead to different results.Source: §5 · p. 9
  • For predictive performance, practitioners might choose RFImportance or CART; F-Test is recommended for interpretability-focused applications.RFImportance and CART are also strong candidates when runtime matters; F-Test ranks highest for stability and validity, with competitive predictive performance and runtime.Source: §5 · p. 9

REPRODUCIBILITY STATEMENT

SelectArena publishes its code and curated datasets and provides experiment and dataset details to support reproducibility.

REPRODUCIBILITY STATEMENT

COMPETING INTERESTS

The authors report that two authors are also authors of TabArena and declare no other competing interests.

COMPETING INTERESTS

A FEATURE SELECTION TAXONOMY FOR PRACTITIONERS

The section complements the traditional structural taxonomy of feature selection with a functional distinction between search and non-search methods. This perspective groups methods by how they approach feature selection rather than where it occurs in a pipeline.

A FEATURE SELECTION TAXONOMY FOR PRACTITIONERS

  • The traditional filter, wrapper, and embedded taxonomy describes whether feature relevance is assessed independently of, through repeated evaluation of, or during training of the downstream model.The section says this structural perspective is less informative to practitioners than categorizing methods by how they approach the optimization problem.Source: A FEATURE SELECTION TAXONOMY FOR… · p. 23
  • Search methods evaluate multiple feature subsets using a model-dependent utility, while non-search methods use proxy mechanisms rather than repeated subset evaluations.Search procedures explore the exponentially large subset space; examples include sequential forward and backward search.Source: A FEATURE SELECTION TAXONOMY FOR… · p. 23
  • Figure A.1 organizes the taxonomy into forward and backward search procedures and non-search groups, extending the latter to regularization-based methods.The non-search grouping builds on Duch (2006), with regularization-based methods associated with Tibshirani (1996) and Zou & Hastie (2005).Source: Figure A.1 · p. 23, A FEATURE SELECTION TAXONOMY FOR… · p. 23

B.1 REVIEWED SURVEYS AND BENCHMARKS

The benchmark’s method-popularity review drew on 33 works, predominantly surveys and benchmarks, and used citation frequency to help determine which methods to exclude.

B.1 REVIEWED SURVEYS AND BENCHMARKS

  • Methods mentioned in fewer than three of the 33 reviewed works were excluded under the method-curation criterion.The review consisted predominantly of surveys and benchmarks.Source: §B.1 · p. 23

B.2 METHOD CURATION CRITERIA

SelectArena curates methods for tabular data and all three task types, while seeking methodological diversity and guarding against applying methods outside their intended settings. Its implementation adapts selected methods to a shared feature-selection pipeline.

B.2 METHOD CURATION CRITERIA

  • Methods were selected for tabular suitability, support for binary, multiclass, and regression tasks, methodological independence, literature recognition, and subgroup coverage.The criteria exclude, for example, task-specific methods and closely related variants.Source: §B.2 · pp. 23–24

B.3 METHOD CURATION INSIGHTS

  • Applying methods beyond their intended target types or data modalities may distort behavior and evaluation, so such uses warrant care.The curation excludes methods intended for categorical targets when their use in regression requires target discretization, and highlights modality-specific assumptions.Source: §B.3 · p. 24
  • Feature-selection research spans largely separate statistical, evolutionary, and XAI streams, which the authors say could benefit from stronger integration.The authors note that these communities make limited use of one another’s methodological advances.Source: §B.3 · p. 24

B.4 METHOD IMPLEMENTATION

  • The implementation integrates feature selectors into AutoGluon’s preprocessing pipeline, with shared cardinality and time constraints and a random-selection fallback for invalid results.Method implementations generally build on scikit-learn or scikit-feature and may return intermediate results when a time limit is reached.Source: §B.4 · p. 25
  • CART and RFImportance score features by mean impurity decrease, while ElasticNet and Lasso use coefficient magnitudes and F-Test ranks features by F-statistic.The implementations use tree-based impurity reductions for CART and forests, joint regularized fits for ElasticNet and Lasso, and top-k F-statistics for F-Test.Source: §B.4.1 · p. 25, §B.4.2 · p. 25, §B.4.3 · p. 25
  • GainRatio divides information gain by feature entropy, Laplacian Score measures locality preservation on a kNN graph, and MI estimates feature-target mutual information.GainRatio’s denominator penalizes high-cardinality features; Laplacian Score uses a graph Laplacian, and MI uses different estimators for discrete and continuous features.Source: §B.4.4 · p. 26, §B.4.5 · p. 26, §B.4.10 · p. 27
  • LOCO ranks features by the proxy-model performance drop on omission, while Markov Blanket removes features whose absence least harms cross-validated performance within a fixed top-k neighborhood.LOCO refits an LGBM proxy once per feature; Markov Blanket uses top-k Pearson-correlated neighbors and adds a regression case to the original formulation.Source: §B.4.7 · p. 26, §B.4.8 · pp. 26–27
  • MRMR adds features by relevance minus average redundancy; Random selects features randomly, while SBE and SFS remove or add features using proxy-model performance.ReliefF and RReliefF compare feature differences across nearest hits and misses; the implementation uses k = 5.Source: §B.4.9 · p. 27, §B.4.11 · p. 27, §B.4.13 · p. 28

B.5 OVERVIEW OF ALL INVESTIGATED FEATURE SELECTION METHODS

The appendices document the investigated-method inventory and dataset-source review, alongside the criteria used to curate datasets for SelectArena.

B.5 OVERVIEW OF ALL INVESTIGATED FEATURE SELECTION METHODS

  • Table B.2 records investigated methods, their years and literature appearances, and inclusion decisions or exclusion reasons.The table also marks some methods as future work when implementations are ongoing.Source: Table B.2 · p. 28, §B.5 · p. 28

C.1 DATASET SOURCES

  • The dataset-source review covers 73 articles screened for real-world tabular datasets.The sources include feature-selection surveys and benchmarks, papers introducing selected methods, and related explainable-AI work.Source: §C.1 · p. 33, Table C.1 · p. 34

C.2 DATASET CURATION CRITERIA

  • Dataset curation requires unique predictive tasks, genuinely tabular data, at least 10 features, real-world origin, IID suitability for random splits, a valid target, and ethical review.Datasets must also be accessible, verifiable, and reliably reusable; inaccessible data are excluded.Source: §C.2 · pp. 34–35

C.3 DATASET CURATION INSIGHTS

The curation insights identify boundaries and risks that affect whether datasets are suitable for feature-selection benchmarking, including modality, feature-space size, split assumptions, and target choice.

C.3 DATASET CURATION INSIGHTS

  • Text- or image-derived data cannot, in the authors’ view, be safely treated as standard tabular data without domain-specific expertise.The paper therefore distinguishes genuine tabular datasets from data derived from other modalities.Source: §C.3 · p. 35
  • Datasets with fewer than 10 features are typically of limited practical relevance for feature selection, and may not support meaningful cardinality budgets.Only a few candidate features are available in such datasets.Source: §C.3 · p. 35
  • Inappropriate splits for temporal or grouped data can cause information leakage and misleading results, while sensitive features warrant critical evaluation.The authors recommend considering both the predictive task and sensitive attributes when judging dataset suitability.Source: §C.3 · p. 35
  • High-dimensional datasets often have limited sample sizes, a combination especially prone to overfitting and overtuning.The paper gives biomedical studies with thousands of genes and few patients as an example.Source: §C.3 · p. 35
  • Determining whether heavily preprocessed, anonymized datasets are real-world or synthetic can be challenging when their original source is no longer identifiable.AutoML challenge data are given as an example.Source: §C.3 · p. 35
  • Scientific discovery datasets may have arbitrarily chosen or artificially constructed targets, producing variants that diverge from the original dataset.The issue arises when datasets not designed for prediction later select targets from existing features.Source: §C.3 · p. 35

C.4 OVERVIEW OF SELECTED DATASETS

The appendices inventory the selected datasets and describe how feature budgets are set by dataset dimensionality. They also outline extended benchmarking results and report that benchmark code and curated datasets are published.

C.4 OVERVIEW OF SELECTED DATASETS

  • The dataset overview lists included datasets with their feature and sample counts, class counts, task types, and usage licenses.The overview includes continued table entries.Source: Table C.2 · pp. 35–38, §C.4 · p. 35

C.5 CARDINALITY CONSTRAINTS

  • For 71 benchmark datasets, cardinality constraints vary by feature dimensionality, with linear spacing for datasets of at most 100 features and exponential spacing above 100.Table C.3 summarizes selected-feature counts by cardinality index.Source: §C.5 · p. 38, Table C.3 · p. 38

D BENCHMARKING DETAILS AND EXTENDED RESULTS

D.1 OPEN-SOURCE BENCHMARKING

  • The authors report that benchmark implementation code and the curated datasets are publicly available.The code is described as under construction.Source: §D.1 · p. 38

D.2 PREDICTIVE PERFORMANCE

This section compares feature-selection methods through rank-based inference and examines tuning and random-fallback effects on predictive performance. It reports budget-dependent ranks, an average tuning lift, and fallback-related shifts that can reflect cardinality violations.

Evaluation metrics

  • Predictive performance uses ROC AUC for binary classification, log-loss for multi-class classification, and RMSE for regression.ROC AUC is prevalence-invariant and threshold-free for comparisons across heterogeneous datasets.Source: §D.2 · pp. 38–39

Elo rating

  • Because methods are static and all matches are available, the benchmark uses Bradley–Terry pairwise ratings, with dataset-cluster bootstrap intervals across 1,000 resamples.The pairwise outcomes compare methods for each downstream-model and budget combination, with scores averaged over seeds.Source: §D.2 · p. 39

Pareto details

  • Pareto analysis compares per-dataset improvement over random and runtime normalized by feature count; errors are averaged across budgets and models.The runtime comparison is reported per 1,000 features.Source: §D.2 · p. 40

Significance analysis

  • With critical distance 2.546, Friedman–Nemenyi found five significant method pairs and all 13 methods beat random and LaplacianScore; RFImportance stayed top-three as budgets grew.CART ranks better at smaller budgets, while F-Test ranks better at larger budgets.Source: §D.2 · p. 41, Figure D.1 · p. 41

Tuning LGBM

  • Tuning LGBM improves the normalized score by 24.7% over the default configuration on average.The comparison averages normalized scores per method and model over datasets and budgets.Source: §D.2 · p. 42, Figure D.2 · p. 42

Random fallback

  • When random fallback was removed, SBE’s +18.05pp effect reflected a cardinality-budget violation, while SFS’s −18.95pp effect accompanied its only meaningful Elo rank drop, from 8 to 13.LOCO’s fallback effect was −1.47pp.Source: §D.2 · pp. 43–44

D.3 STABILITY

F-Test and distance-based methods consistently rank highly for stability, with F-Test especially strong at very low events per variable (EPV). Lower EPV narrows the set of relatively stable methods, while most methods show intermediate-to-good stability overall.

D.3 STABILITY

  • F-Test, LaplacianScore, MI, GainRatio, (R)ReliefF, and mRMR consistently rank highest in stability across EPVs.The near-optimal set within 3% of the best-performing method is dominated by the same group.Source: §D.3 · p. 44, Figure D.4 · p. 45, Figure D.5 · p. 45
  • F-Test ranks highest at very low EPVs, where other methods are rarely within 3% of its stability.It also often performs well at high and medium EPVs.Source: §D.3 · p. 44, Figure D.5 · p. 45
  • At high and medium EPVs, a broader diversity of methods achieves competitive stability than at very low EPVs.At very low EPVs, only a few methods are relatively stable.Source: §D.3 · p. 44, Figure D.4 · p. 45
  • Lasso, SBE, SFS, and LOCO have low stability, while the remaining methods show intermediate-to-good stability.The reported timeouts are 46.41% for SBE, 45.28% for SFS, and 23.10% for LOCO; stability below 0.40 is classified as poor, and 0.40–0.75 as intermediate-to-good.Source: §D.3 · p. 44

D.4 VALIDITY

F-Test and information-theoretic methods dominate validity rankings across noise levels, followed by LaplacianScore. Increasing noise lowers validity scores across methods and makes differences between high- and low-ranked methods more pronounced.

D.4 VALIDITY

  • F-Test and information-theoretic methods dominate validity rankings, followed by LaplacianScore, consistently across noise levels.The reported comparisons cover 50%, 75%, and 100% added noise.Source: §D.4 · p. 46, Figure D.6 · p. 47, Figure D.7 · p. 47
  • Every method’s validity score declines as noise increases; the table reports mean validity and stability across datasets.The validity evaluation compares noise levels of 50%, 75%, and 100% of the original feature count.Source: §D.4 · p. 46, Table D.4 · p. 46, Figure D.6 · p. 47
  • F-Test’s validity drops by less than 4% per noise increment, compared with MarkovBlanket’s 9% decline per step.Method rankings change only slightly across noise levels.Source: §D.4 · p. 46, Figure D.6 · p. 47, Figure D.8 · p. 48
  • As noise increases, fewer low-ranked methods remain within 3% of the best method, accentuating differences between high- and low-ranked methods.The thresholded comparisons are shown at 50%, 75%, and 100% added noise.Source: §D.4 · p. 46, Figure D.7 · p. 47, Figure D.9 · p. 48

D.5 RAW PREDICTIVE PERFORMANCE

This section compiles raw predictive-performance results across downstream models, including per-cell values grouped by metric where specified and reported with ± quantities, often as three-seed means and sample standard deviations. Several supplied passages omit method, task, or metric labels, limiting attribution of individual values.

D.5 RAW PREDICTIVE PERFORMANCE

  • Metric errors are reported per method across datasets and budgets, aggregated over three seeds.Results are split across tables for LGBM, Linear model, RF, RealMLP, and TabICLv2.Source: §D.5 · p. 50
  • The LGBM table reports per-cell means over three seeds, with sample standard deviations shown as subscripts.The supplied table values lack row and column labels, so they do not support identifying specific method comparisons.Source: Table D.5 · p. 50
  • The supplied passages contain raw numeric result sequences but provide no row or column labels for interpreting individual values.Source: §D.5 · p. 53
  • Table D.6 reports per-cell results using the downstream Linear model.The table groups tasks by metric.Source: Table D.6 · p. 55
  • Each cell reports the mean over three seeds and the sample standard deviation as a subscript.Source: Table D.6 · p. 55
  • The unlabeled numeric entries do not identify their methods, datasets, or metrics, so no performance ranking can be supported from this excerpt.The passages show values with ± terms but provide no table headers or explanatory prose.Source: §D.5 · p. 56
  • Table D.7 reports per-cell downstream-RF results grouped by metric, with each cell showing the mean over three seeds and sample standard deviation as a subscript.Source: Table D.7 · p. 60
  • The opening rows contain unlabeled values such as .796 ±.047 and .754 ±.063.The supplied text gives no metric or method labels for interpreting these entries.Source: §D.5 · p. 61
  • Further rows continue the same unlabeled numeric format, including 1.38 ±.14 and 1.01 ±.04.Without accompanying headers, the values cannot be assigned to specific methods or tasks from these passages.Source: §D.5 · p. 61
  • Later entries likewise provide decimal values with ± terms, including .717 ±.068 and .624 ±.007.The supplied passages do not identify what methods or tasks these values represent.Source: §D.5 · p. 61
  • The extracted score sequences lack method and metric labels, so they do not establish which method performs best.The passages provide numeric values with ± terms but no corresponding column headers.Source: §D.5 · p. 63
  • Score sequences are shown for online_shoppers_purchasing and contaminant_detection_10GHz, but their values cannot be assigned to methods from these passages.The dataset names appear adjacent to numeric sequences without method labels.Source: §D.5 · pp. 64–65
  • The supplied score sequences have no accompanying row or column labels, so individual values cannot be assigned to methods or metrics.Source: §D.5 · p. 65
  • Without those labels, this extract does not establish which method performs best or support metric-specific comparisons.Source: §D.5 · p. 65
  • Table D.8 reports per-cell downstream RealMLP results, grouped by metric, with means over three seeds and sample standard deviations as subscripts.Source: Table D.8 · p. 66
  • The excerpt contains many mean ± spread values, but without retained method or metric labels it cannot establish which method performs best.The numeric sequences are presented without the row and column headings needed to interpret them.Source: §D.5 · p. 69
  • The visible dataset labels include online_shoppers_purchasing and contaminant_detection_10GHz.Source: §D.5 · p. 70
  • Table D.9 reports per-cell results for downstream model TabICLv2, grouped by metric, with means over three seeds and sample standard deviations.The supplied passages do not include row or column labels, so the individual scores cannot be assigned to specific methods or tasks.Source: Table D.9 · p. 71
  • The results are displayed as repeated score ± quantity entries across many passages.The supplied text does not identify what the ± quantities represent.Source: §D.5 · p. 74
  • The visible numerical entries span lower values around 0.5 and values reaching 1.00, but their method and task associations are not shown.For example, the supplied entries include .490 ±.092 and 1.00 ±.00.Source: §D.5 · p. 75
  • The provided passages lack table headers, so they do not support identifying a best-performing method or comparing specific methods.The numerical sequences cannot be mapped to method, metric, or task labels from this text alone.Source: §D.5 · p. 75
  • 0.596 ±.004 to 0.635 ±.001: entries in one result passage span this range.Source: §D.5 · p. 76
  • 0.572 ±.121 to 0.838 ±.057: entries in another result passage span this range.Source: §D.5 · p. 76
  • 0.455 ±.046 to 0.988 ±.001: the supplied score entries span this range.These endpoints occur in separate result passages; the excerpt does not identify their methods or tasks.Source: §D.5 · p. 76
Loading 2610.13644v1…