Divide, Specialize, and Route: A New Approach to Efficient Ensemble Learning

Anna Kozak; Antoni Zajko; Dawid P{\l}udowski; Jakub Piwko; J\k{e}drzej Ruci\'nski; Katarzyna Wo\'znica; Mateusz Zacharecki; Patryzja \.Zak

REVIEW 4 major objections 5 minor 21 references

Divide, Specialize, and Route: A New Approach to Efficient Ensemble Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Hellsemble builds ensembles by partitioning data into difficulty circles and routing each instance to a specialist model.

desk verdict A promising routing-based ensemble idea that is undermined by its own selective evaluation: the reported gains only hold on runs where Hellsemble formed multiple models, and classical ensemble baselines are never tested. read the letter →

arxiv 2506.20814 v1 pith:I6CHOLKD submitted 2025-06-25 cs.LG

Jakub Piwko , Jędrzej Ruciński , Dawid P{l}udowski , Antoni Zajko , Patryzja Żak , Mateusz Zacharecki , Anna Kozak , Katarzyna Woźnica This is my paper

classification cs.LG

keywords ensemblelearningdynamicselectionbinaryclassificationmodelroutingdatasetcomplexityinterpretabilitytabulardata

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Hellsemble, a framework for binary classification, builds its committee by passing instances that one model misclassifies to the next model, so each base learner trains on a harder slice of the data. The paper calls these slices 'circles of difficulty.' A separate router is trained to map each new instance to the circle—and therefore to the base model—best suited to it. The central claim is that this divide-specialize-route design lets simple classifiers match or beat stronger global models, and the reported experiments show gains of 2–3 percentage points in average accuracy when the base suite contains simple models. The significance is practical: ensembles usually pay for their accuracy with training and inference cost, while Hellsemble needs only the router plus one specialist at prediction time.

What carries the argument

The load-bearing object is the circle of difficulty: the subset of training data handed to each successive base model, defined by the previous model's errors. The circle labels double as training targets for the router, so the same partition that specializes the base models also creates the routing signal. In the greedy variant, the next model is chosen by its validation-score improvement, and training stops when no candidate improves the ensemble, which is what keeps the final system small and cheap.

What would settle it

Shuffle the circle-of-difficulty labels before training the router and rerun the benchmark suite; if test accuracy is unchanged, the router is not using difficulty structure, and the committee alone accounts for any gains.

Watch

Extended reading notes

Core claim

The paper's central discovery is that treating instance-level difficulty as an explicit training signal can replace the usual ensemble aggregation step. Hellsemble starts by training a base model on the full training set, then removes the instances that model got right and trains the next base model on the remaining 'harder' examples. Repeating this yields a chain of models ordered by difficulty, and the iteration index at which an instance is filtered out becomes its circle label. A router trained as a multiclass classifier on those labels predicts, for a new instance, which chain element should answer. On the OpenML-CC18 and Tabzilla benchmarks, the authors report that this arrangement outperforms the individual base models in most tested configurations, with the clearest gains—2 to 3 percentage points—when the base models are simple classifiers.

Load-bearing premise

The router must learn a stable mapping from features to difficulty circles that holds on new data; if test instances fall into different difficulty patterns than the training set, the router sends them to the wrong specialist and the ensemble loses its advantage.

Editorial extensions

If this is right

With simple base models, Hellsemble reports average accuracy gains of 2–3 percentage points over each individual base model.
At inference time only the router and one selected base model run, so prediction cost stays near that of a single classifier.
When no additional model improves validation score, Hellsemble stops and returns a single model, so it does not force ensemble overhead.
The circle labels give an interpretable segmentation of the data, showing which instances are easy or hard for each learner.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The router's labels come from training-set errors and are noisy; confidence-weighted or iteratively re-labeled circles could improve routing accuracy beyond the paper's reported results.
The difficulty circles resemble a curriculum, so choosing the order of models by complementary errors rather than validation accuracy alone might create larger gains.
The reported gains are strongest when the base suite lacks a powerful global model, suggesting the method is a way to upgrade weak learners rather than to beat strong ones.
The router is itself a classifier, so its capacity may become the bottleneck; using a stronger router than the base models could turn the 2–3 point gains into larger margins.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The core idea is worth knowing: train a sequence of simple models on increasingly difficult subsets, then train a router to pick which specialist handles each test instance. That's a clean, interpretable way to get ensemble diversity with an inference cost of one model plus a router. I agree with the reader that this is a legitimate new combination, even though each ingredient—boosting-style sequential focus, DES, query routing—has precedent.

What the paper does well: the algorithm is clearly described, the sequential versus greedy trade-off is sensible, and the authors are upfront about when Hellsemble collapses to a single model. Figure 2, showing how often multi-model ensembles form, is actually useful because it lets you see the selection issue directly.

The problem is the evaluation. The abstract says Hellsemble “often outperforms classical ensemble methods,” but no bagging, boosting, or DES baselines appear anywhere in the experiments. Every comparison is against individual base models. And the accuracy analysis in Section 5 is restricted to runs where Hellsemble formed more than one model. Since the stopping rule is based on validation improvement, single-model runs are exactly the cases where Hellsemble could not beat its best base model. Dropping them inflates the average. Figure 2 shows most configurations are majority single-model, so the filtered subset is not representative. The 2–3 percentage point improvement is therefore conditional on a favorable selection event, not on the procedure as a whole.

Router generalization is a real concern, but secondary. Even a perfectly generalizing router would not fix the missing baselines or the filtering. On the positive side, the authors do not hide the restriction—they state it plainly—but stating it does not make the claim unsubstantiated. Also missing: error bars, per-dataset counts, and code. For a method whose practical value depends on how often it actually helps, that is a serious omission.

Who is this for? Someone working on efficient ensembles or dynamic selection might find the idea worth a look. It deserves a serious referee because the algorithmic idea is coherent and the flaws are fixable with a proper comparison and full reporting. I would not cite it in its current form, but I would bring it to a reading group as a case study in evaluation selection bias.

Recommendation: send it to peer review, but with a clear request for (a) bagging/boosting/DES baselines, (b) results on all runs, not just multi-model ones, and (c) error bars or per-dataset numbers. If those do not come, reject.

Referee Report

4 major / 5 minor

Summary. The paper proposes Hellsemble, an ensemble method for binary classification that builds a committee by iteratively training base models on increasingly difficult subsets of the training data: after each model is added, misclassified instances (plus a fraction α of correctly classified instances) are passed to the next round, and a router classifier is trained on the resulting 'circle of difficulty' labels to choose a single base model at test time. Two training procedures are given (sequential, Algorithm 1, and greedy, Algorithm 2), and experiments are run on OpenML-CC18 and Tabzilla with 16 combinations of base-model suites and router models, comparing against individual base models. The paper claims that Hellsemble often outperforms both its base models and, in the abstract, classical ensemble methods, while being efficient and interpretable.

Significance. If the claims were supported, the Hellsemble idea would be a useful contribution: it provides a simple, interpretable way to specialize simple learners to subregions of the input space and routes at inference through a single model, which could indeed reduce inference cost relative to full ensembles. The framework's core loop in Algorithm 1 is clearly described and easy to implement. However, the empirical evidence in the manuscript does not substantiate the abstract's headline claims: comparisons to classical ensemble methods are absent, and the accuracy analysis is restricted to a selected subset of runs. The paper also ships no code and reports no uncertainty quantification, so the central quantitative claims are not currently verifiable.

major comments (4)

[Section 5, Figure 2, Algorithm 1] The accuracy analysis restricts attention to 'dataset and experiment pairs where the resulting Hellsemble (either sequential or greedy) included more than one model' (Section 5). Since Figure 2 shows that single-model Hellsembles are the majority in most configurations, this filtering excludes the majority of runs and removes exactly the cases where Hellsemble cannot beat its best base model. Algorithm 1 stops adding models when the validation score does not improve (lines 12–13), so multi-model formation is itself a favorable selection event. The reported 2–3 percentage point improvements therefore apply only to a non-representative subset, and the claim that Hellsemble 'often outperforms' its base models is not supported by the full procedure.
[Abstract and Section 4] The abstract's claim that Hellsemble 'often outperforms classical ensemble methods' is never tested. The experimental protocol compares Hellsemble only with the individual base models in each suite; there are no bagging, boosting, stacking, or dynamic ensemble selection baselines, nor comparisons to standard ensemble implementations. Consequently, the central comparative claim is unsubstantiated as written.
[Section 5, Figures 3–6] The accuracy plots report only averages over datasets, without error bars, per-dataset counts, or statistical tests. Given that the number of datasets forming multi-model ensembles varies by configuration, the reader cannot distinguish a consistent improvement from an artifact of averaging over few favorable datasets. The manuscript should report full per-dataset results, win/loss counts, and appropriate significance measures.
[Section 3, Algorithm 1 (router training, lines 8–10)] The method depends on the router's ability to learn a stable mapping from features to 'circle of difficulty' labels derived from training-set errors, but no analysis of router generalization is provided. The authors should report router accuracy versus an oracle router (which would always pick the model with the best test loss) and versus random routing, and should discuss the failure mode in which misrouting sends an instance to a specialist that is worse than the best base model. A sensitivity analysis for the α parameter and the validation-based stopping threshold would also clarify which components of the algorithm drive performance.

minor comments (5)

[Figure 1 and Section 5] Figure 1's caption contains a typo ('evalution'), and Section 5 refers to 'configurations 3 and 4' although Table 1 does not number the configurations; please clarify which configurations are meant.
[References] References [12] and [13] point to the same AutoGluon paper; the duplicate should be removed or replaced with the relevant distinct citation.
[Acknowledgments and Section 4] The acknowledgments contain a typo ('programme.program'), and the phrase 'a robust suites' should be 'a robust suite'.
[Section 4] No random seeds, exact dataset versions, or code repository are provided, which makes the experimental results difficult to reproduce.
[Section 3.1, Algorithm 1] The hyperparameter α (the fraction of correctly classified instances passed forward) and the validation threshold are not specified for the experiments; please report the values used and any sensitivity checks.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is an empirical procedure evaluated on held-out data, not a derivation that reduces to its own inputs.

full rationale

The paper proposes Hellsemble, an iterative ensemble procedure in which base models are trained on subsets defined by previous models' misclassifications, and a router is trained on instance-level labels derived from the iteration at which each training instance was filtered out. This is a constructive algorithm, not a mathematical derivation: there is no equation in which an output is defined in terms of the claimed result, and no fitted parameter is renamed as a prediction. The router is trained on training-set error labels and evaluated on held-out test data, which is standard practice rather than circular reasoning. The paper does cite prior work, but the cited references are standard benchmarks and related methods, not a self-citation chain that carries the load of the central claim. The main weaknesses are empirical rather than circular: the accuracy analysis in Section 5 is restricted to runs where Hellsemble formed a multi-model ensemble, and the abstract's claim that Hellsemble 'often outperforms classical ensemble methods' is not directly tested against bagging, boosting, or DES in the reported experiments. These are important correctness and reporting concerns, but they do not constitute a reduction of the claimed result to its own inputs. Therefore the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method rests on several domain assumptions about the stability and learnability of instance difficulty, plus a set of hand-chosen design parameters (alpha, stopping rule, router and base model selections). No formal derivation or external benchmark is provided, so the empirical claim stands on the experimental setup, which is selective and lacks standard ensemble baselines.

free parameters (4)

alpha (fraction of correctly classified instances carried to next iteration) = Not specified
Controls overlap between difficulty circles; the paper says it is 'determined based on the performance difference' (Section 3) but gives no concrete value or procedure.
Validation-based stopping threshold
The stopping rule in Algorithm 1 is Scurr <= Sprev, with no tolerance. This directly determines how many models are included and heavily influences the reported accuracy, which only covers multi-model runs.
Router model choice and hyperparameters
Results in Figures 3 to 6 vary strongly across the 4 router models and 4 base model suites; averages are reported over these manually selected configurations, so the headline numbers depend on this choice.
Base model suite composition
The paper tests 4 curated sets of base models; the conclusion that Hellsemble helps simple models is derived from this specific curation.

assumptions (4)

domain assumption Misclassified instances are intrinsically harder and can be learned by later models
The entire 'circles of difficulty' construction in Section 3 assumes error patterns are transferable and stable across models, and that a later model can learn what an earlier one missed.
domain assumption The router can learn a generalizable mapping from features to difficulty circles
The router is trained on labels derived from training-set errors; the paper assumes these labels are predictable from features on test data, but provides no analysis of this assumption.
domain assumption Validation accuracy with the current router is a reliable criterion for model selection and stopping
Algorithm 1 uses Scurr, the validation score of the ensemble routed by the freshly trained router, to decide whether to keep adding models. This assumes validation performance is a stable proxy for test performance, which is standard but unstated and especially fragile when the router changes each iteration.
ad hoc to paper The fraction alpha of correctly classified instances passed forward is sufficient regularization
The paper states alpha is chosen to minimize the gap between training and validation performance, but no formula, search, or default is given, so this is effectively an ad hoc design choice.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Divide, Specialize, and Route: A New Approach to Efficient Ensemble Learning." pith.science (2026). https://pith.science/paper/I6CHOLKD

@misc{pith2026250620814,
  author       = {Pith},
  title        = {Pith review of: Divide, Specialize, and Route: A New Approach to Efficient Ensemble Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I6CHOLKD}},
  note         = {Machine review of arXiv:2506.20814}
}

read the original abstract

Ensemble learning has proven effective in boosting predictive performance, but traditional methods such as bagging, boosting, and dynamic ensemble selection (DES) suffer from high computational cost and limited adaptability to heterogeneous data distributions. To address these limitations, we propose Hellsemble, a novel and interpretable ensemble framework for binary classification that leverages dataset complexity during both training and inference. Hellsemble incrementally partitions the dataset into circles of difficulty by iteratively passing misclassified instances from simpler models to subsequent ones, forming a committee of specialised base learners. Each model is trained on increasingly challenging subsets, while a separate router model learns to assign new instances to the most suitable base model based on inferred difficulty. Hellsemble achieves strong classification accuracy while maintaining computational efficiency and interpretability. Experimental results on OpenML-CC18 and Tabzilla benchmarks demonstrate that Hellsemble often outperforms classical ensemble methods. Our findings suggest that embracing instance-level difficulty offers a promising direction for constructing efficient and robust ensemble systems.

Figures

Figures reproduced from arXiv: 2506.20814 by the authors.

**Figure 1.** Training and dataset evalution schema for Hellsemble The training procedure and evolution of the dataset across iterations are illustrated in [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

**Figure 2.** Proportion of experiments in each configuration where the Hellsemble (either sequential or greedy) consisted of exactly one model versus more than one model. Before analyzing performance metrics, we first examine how frequently Hellsemble constructs ensembles consisting of more than one model. Importantly, we did not enforce multi-model construction during training—Hellsemble naturally stops adding models when the … view at source ↗

**Figure 3.** Average accuracy across all datasets for different routing models and the first base model suite (e.g., KNN, Naive Bayes, Logistic Regression, and Decision Trees). in terms of average accuracy. The one exception is when using KNN with default settings as the router—here, the Decision Tree performs best. Nevertheless, these results indicate that even with basic classifiers, Hellsemble can yield accuracy improvements … view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Average accuracy across all datasets for different routing models and the second base model suite (e.g., Random Forests and XGBoost) [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

**Figure 5.** Figure 5: Average accuracy across all datasets for different routing models and the third base model suite (composed of simpler classifiers). In [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

**Figure 6.** Figure 6: Average accuracy across all datasets for different routing models and the fourth base model suite (containing both simple and advanced models). Finally, [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 12 canonical work pages

[1]

Pervasive and Mobile Computing 104, 101973 (2024)

Bekku, H., Kume, T., Tsuge, A., Nakazawa, J.: A stable and efficient dynamic ensemble method for pothole detection. Pervasive and Mobile Computing 104, 101973 (2024)

work page 2024
[2]

Bischl, B., Casalicchio, G., Feurer, M., Gijsbers, P., Hutter, F., Lang, M., Manto- vani, R.G., van Rijn, J.N., Vanschoren, J.: Openml benchmarking suites (2021), https://arxiv.org/abs/1708.03731

arXiv 2021
[3]

Machine learning45, 5–32 (2001)

Breiman, L.: Random forests. Machine learning45, 5–32 (2001)

2001
[4]

In: Sixth International Conference on Data Mining (ICDM’06)

Caruana, R., Munson, A., Niculescu-Mizil, A.: Getting the most out of ensemble selection. In: Sixth International Conference on Data Mining (ICDM’06). pp. 828–

work page
[5]

In: Proceedings of the twenty-first international conference on Machine learning

Caruana, R., Niculescu-Mizil, A., Crew, G., Ksikes, A.: Ensemble selection from libraries of models. In: Proceedings of the twenty-first international conference on Machine learning. p. 18 (2004)

work page 2004
[6]

Neural computing and applications22, 673–688 (2013)

Cavalin, P.R., Sabourin, R., Suen, C.Y.: Dynamic selection approaches for multiple classifier systems. Neural computing and applications22, 673–688 (2013)

work page 2013
[7]

In: Proceedings of the 22nd ACM SIGKDD International Con- ference on Knowledge Discovery and Data Mining

Chen, T., Guestrin, C.: Xgboost: A scalable tree boosting sys- tem. In: Proceedings of the 22nd ACM SIGKDD International Con- ference on Knowledge Discovery and Data Mining. p. 785–794. KDD ’16, ACM (Aug 2016). https://doi.org/10.1145/2939672.2939785, http://dx.doi.org/10.1145/2939672.2939785

arXiv 2016
[8]

Pattern recognition48(5), 1925–1935 (2015)

Cruz,R.M.,Sabourin,R.,Cavalcanti,G.D.,Ren,T.I.:Meta-des:Adynamicensem- ble selection framework using meta-learning. Pattern recognition48(5), 1925–1935 (2015)

work page 2015

Show all 21 references

[9]

Ensemble machine learn- ing: Methods and applications pp

Cutler, A., Cutler, D.R., Stevens, J.R.: Random forests. Ensemble machine learn- ing: Methods and applications pp. 157–175 (2012)

2012
[10]

In: International work- shop on multiple classifier systems

Dietterich, T.G.: Ensemble methods in machine learning. In: International work- shop on multiple classifier systems. pp. 1–15. Springer (2000)

2000
[11]

arXiv preprint arXiv:2404.14618 (2024)

Ding, D., Mallick, A., Wang, C., Sim, R., Mukherjee, S., Ruhle, V., Lakshmanan, L.V., Awadallah, A.H.: Hybrid llm: Cost-efficient and quality-aware query routing. arXiv preprint arXiv:2404.14618 (2024)

2024 arXiv
[13]

Erickson, N., Mueller, J., Shirkov, A., Zhang, H., Larroy, P., Li, M., Smola, A.: Autogluon-tabular: Robust and accurate automl for structured data (2020), https://arxiv.org/abs/2003.06505

2020 arXiv
[14]

Journal of Machine Learning Research 23(261), 1–61 (2022)

Feurer, M., Eggensperger, K., Falkner, S., Lindauer, M., Hutter, F.: Auto-sklearn 2.0: Hands-free automl via meta-learning. Journal of Machine Learning Research 23(261), 1–61 (2022)

2022
[15]

Journal of computer and system sciences55(1), 119–139 (1997)

Freund, Y., Schapire, R.E.: A decision-theoretic generalization of on-line learning and an application to boosting. Journal of computer and system sciences55(1), 119–139 (1997)

1997
[16]

Pattern Recognition34(9), 1879–1881 (2001)

Giacinto, G., Roli, F.: Dynamic classifier selection based on multiple classifier be- haviour. Pattern Recognition34(9), 1879–1881 (2001)

2001
[17]

Pattern recognition41(5), 1718–1731 (2008)

Ko, A.H., Sabourin, R., Britto Jr, A.S.: From dynamic classifier selection to dy- namic ensemble selection. Pattern recognition41(5), 1718–1731 (2008)

2008
[18]

McElfresh,D.,Khandagale,S.,Valverde,J.,C,V.P.,Feuer,B.,Hegde,C.,Ramakr- ishnan, G., Goldblum, M., White, C.: When do neural nets outperform boosted trees on tabular data? (2024), https://arxiv.org/abs/2305.02997 14 Piwko, Ruciński, Płudowski, Zajko, Żak, Zacharecki

2024 arXiv
[19]

In: International Conference on Machine Learning

McGill, M., Perona, P.: Deciding how to decide: Dynamic routing in artificial neu- ral networks. In: International Conference on Machine Learning. pp. 2363–2372. PMLR (2017)

2017
[20]

Olson, R.S., Bartley, N., Urbanowicz, R.J., Moore, J.H.: Evaluation of a tree-based pipeline optimization tool for automating data science (2016), https://arxiv.org/abs/1603.06212

2016 arXiv
[21]

Wiley interdisciplinary reviews: data min- ing and knowledge discovery5(1), 21–34 (2015)

Sesmero, M.P., Ledezma, A.I., Sanchis, A.: Generating ensembles of heterogeneous classifiers using stacked generalization. Wiley interdisciplinary reviews: data min- ing and knowledge discovery5(1), 21–34 (2015)

2015
[22]

IEEE transactions on pattern analysis and machine intelligence 19(4), 405–410 (1997)

Woods, K., Kegelmeyer, W.P., Bowyer, K.: Combination of multiple classifiers using local accuracy estimates. IEEE transactions on pattern analysis and machine intelligence 19(4), 405–410 (1997)

1997

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Pervasive and Mobile Computing 104, 101973 (2024)

Bekku, H., Kume, T., Tsuge, A., Nakazawa, J.: A stable and efficient dynamic ensemble method for pothole detection. Pervasive and Mobile Computing 104, 101973 (2024)

work page 2024

[2] [2]

Bischl, B., Casalicchio, G., Feurer, M., Gijsbers, P., Hutter, F., Lang, M., Manto- vani, R.G., van Rijn, J.N., Vanschoren, J.: Openml benchmarking suites (2021), https://arxiv.org/abs/1708.03731

arXiv 2021

[3] [3]

Machine learning45, 5–32 (2001)

Breiman, L.: Random forests. Machine learning45, 5–32 (2001)

2001

[4] [4]

In: Sixth International Conference on Data Mining (ICDM’06)

Caruana, R., Munson, A., Niculescu-Mizil, A.: Getting the most out of ensemble selection. In: Sixth International Conference on Data Mining (ICDM’06). pp. 828–

work page

[5] [5]

In: Proceedings of the twenty-first international conference on Machine learning

Caruana, R., Niculescu-Mizil, A., Crew, G., Ksikes, A.: Ensemble selection from libraries of models. In: Proceedings of the twenty-first international conference on Machine learning. p. 18 (2004)

work page 2004

[6] [6]

Neural computing and applications22, 673–688 (2013)

Cavalin, P.R., Sabourin, R., Suen, C.Y.: Dynamic selection approaches for multiple classifier systems. Neural computing and applications22, 673–688 (2013)

work page 2013

[7] [7]

In: Proceedings of the 22nd ACM SIGKDD International Con- ference on Knowledge Discovery and Data Mining

Chen, T., Guestrin, C.: Xgboost: A scalable tree boosting sys- tem. In: Proceedings of the 22nd ACM SIGKDD International Con- ference on Knowledge Discovery and Data Mining. p. 785–794. KDD ’16, ACM (Aug 2016). https://doi.org/10.1145/2939672.2939785, http://dx.doi.org/10.1145/2939672.2939785

arXiv 2016

[8] [8]

Pattern recognition48(5), 1925–1935 (2015)

Cruz,R.M.,Sabourin,R.,Cavalcanti,G.D.,Ren,T.I.:Meta-des:Adynamicensem- ble selection framework using meta-learning. Pattern recognition48(5), 1925–1935 (2015)

work page 2015