SPOCC: Scalable POssibilistic Classifier Combination -- toward robust aggregation of classifiers

John Klein; Mahmoud Albardan; Olivier Colot

REVIEW 3 major objections 4 minor 52 references

SPOCC: Scalable POssibilistic Classifier Combination -- toward robust aggregation of classifiers

T0 review · 3 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a possibility-theoretic classifier combination method, SPOCC/adaSPOCC, is asymptotically robust to random-guess, adversarial, and duplicate classifiers while remaining scalable and incremental.

desk verdict A genuinely novel way to fuse label-only classifiers using possibility theory, with careful experiments, but the proof sketches do not support the headline robustness guarantees for multiclass problems. read the letter →

arxiv 1908.06475 v2 pith:46WIZKR3 submitted 2019-08-18 cs.LG stat.ML

Mahmoud Albardan , John Klein , Olivier Colot This is my paper · ORCID

classification cs.LGstat.ML

keywords robustclassifiercombinationagnosticaggregationpossibilitytheoryinformationfusiont-normensembleadversarialrobustnessincremental

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

This paper aims to establish that classifier ensembles can be made robust by converting each classifier's label predictions into possibility distributions and combining them with an adaptive t-norm. The proposed method, SPOCC, and its adaptive variant adaSPOCC, prove that asymptotically, a random-guess classifier has no influence, an adversarial classifier can be rectified without changing the ensemble output, and exact copies of a classifier add no weight. These properties would matter for decentralized learning, where black-box classifiers trained on separate data must be aggregated without retraining. The paper also claims scalability, with parameters linear in the number of classifiers, and incremental addition of new classifiers.

What carries the argument

The central object is the possibilistic ensemble: each classifier's validation confusion matrix is normalized into conditional probabilities $p(Y\mid c_k=\omega_j)$, then converted by the Dubois-Prade transform, the unique probability-to-possibility transformation that preserves consistency, preference order, and maximal specificity, into a possibility distribution $\pi_{k\mid j}$. Predictions are combined elementwise by Aczel-Alsina t-norms $T_\lambda(a,b)=\exp\left(-\left(|\log a|^\lambda+|\log b|^\lambda\right)^{1/\lambda}\right)$, which interpolate between the product t-norm at $\lambda=1$ and the idempotent minimum at $\lambda=\infty$; the minimum is what neutralizes duplicate classifiers. adaSPOCC builds a binary dendrogram via hierarchical agglomerative clustering on the dependence measure $\kappa(c_k,c_{k'})=1-\exp\left(-\frac{1}{n_{\mathrm{val}}}|\log(L_0/L_1)|\right)$, and discounts weak classifiers using $\alpha_k=1-\left(\frac{1-\hat{r}[c_k]}{1-\min_{k'}\hat{r}[c_{k'}]}\right)^\rho$. These pieces together carry the robustness argument.

What would settle it

Take a fixed ensemble containing two exact duplicate classifiers and a validation set; compute adaSPOCC's aggregate output with both copies and after removing one, and check whether the outputs are identical as the validation set grows. For the adversarial property, construct a classifier with known worse-than-random error and test whether replacing its predictions with the posterior-mode rectification leaves the ensemble output unchanged on every test input; a single differing input at large validation-set size would falsify the claim.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the three robustness properties (a) through (c) hold asymptotically for adaSPOCC: a classifier whose error rate equals random guess is neutralized, because its conditional distributions become uniform and the probability-to-possibility transform maps them to the constant-one possibility distribution, which is the neutral element of every t-norm used; a classifier with worse-than-random error rate can be rectified through the posterior mode map $c_{\mathrm{rec}}(x)=\arg\max_y p(Y=y\mid c_k(x))$, and asymptotically the ensemble output is unchanged; and exact copies are detected through the likelihood-ratio dependence measure, grouped by hierarchical agglomerative clustering, and merged with an idempotent minimum t-norm. The paper argues these follow from almost-sure convergence of multinomial estimates on the validation set, so they hold systematically for large validation sets rather than only in expectation or with high probability.

Load-bearing premise

The load-bearing premise is that the automatic tuning step actually selects the most cautious combination rule for groups of dependent or duplicate classifiers, and that the rectified adversarial classifier's error pattern matches the original in the way the proof requires; if either fails, the claimed guarantees can fail.

Editorial extensions

If this is right

If a random-guess classifier is added to an ensemble, adaSPOCC's aggregate prediction is asymptotically the same as without it, so fault tolerance does not require explicit failure detection.
If a malicious or broken classifier has above-random error, mapping its predictions through the posterior mode rectifier leaves the ensemble output unchanged, neutralizing the adversary's influence.
Duplicate classifiers do not bias the aggregate toward their shared prediction, so vote-based majority distortions from correlated or redundant ensemble members are avoided.
The parameter count is $O(\ell^2K)$, linear in the number of classifiers, and new classifiers can be appended by computing only their own confusion matrix and updating the cluster structure, which suits decentralized settings.
In the reported experiments, adaSPOCC achieved the highest average rank across eight real datasets and the smallest maximal accuracy shortfall relative to the best competing method, at 2.25%.
If a random-guess classifier is added to an ensemble, adaSPOCC's aggregate prediction is asymptotically the same as without it, so fault tolerance does not require explicit failure detection.
If a malicious or broken classifier has above-random error, mapping its predictions through the posterior mode rectifier leaves the ensemble output unchanged, neutralizing the adversary's influence.
Duplicate classifiers do not bias the aggregate toward their shared prediction, so vote-based majority distortions from correlated or redundant ensemble members are avoided.

Reading between the lines

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

The same machinery suggests a continuous version of the duplicate-neutrality property: exact copies are the extreme case, while increasingly correlated classifiers receive increasingly strong idempotent combination through larger $\lambda$ values.
The paper's neutralization argument depends on the Dubois-Prade transform being the unique specificity-maximizing probability-to-possibility transform; a different transform would change the neutrality of uniform distributions, so this choice is load-bearing beyond convenience.
A testable extension would measure how the three robustness properties degrade as the validation set shrinks, giving practitioners a concrete sense of the asymptotic threshold in finite-sample regimes.
Because the method is agnostic to base learners and uses only label predictions, it could extend naturally to settings where classifiers are trained locally and only their predictions on a shared validation set are exposed.
The same machinery suggests a continuous version of the duplicate-neutrality property: exact copies are the extreme case, while increasingly correlated classifiers receive increasingly strong idempotent combination through larger $\lambda$ values.
The paper's neutralization argument depends on the Dubois-Prade transform being the unique specificity-maximizing probability-to-possibility transform; a different transform would change the neutrality of uniform distributions, so this choice is load-bearing beyond convenience.
A testable extension would measure how the three robustness properties degrade as the validation set shrinks, giving practitioners a concrete sense of the asymptotic threshold in finite-sample regimes.
Because the method is agnostic to base learners and uses only label predictions, it could extend naturally to settings where classifiers are trained locally and only their predictions on a shared validation set are exposed.

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

You should read the experiments and the construction; the proofs in Section 3.7 are not in a shape that supports the headline claims. The paper introduces SPOCC/adaSPOCC, which turns confusion-matrix columns into possibility distributions via the Dubois-Prade transform, then aggregates with Aczel-Alsina t-norms whose parameters are set adaptively via HAC clustering and discounting. That's new for classifier combination, as far as I know, and the formulation of robustness properties (a)-(c) is a useful step.

What's good: the method is genuinely agnostic to base learners, has linear parameter growth in K, supports incremental updates, and the experiments are thorough. The synthetic and real data comparisons include many baselines and bootstrap confidence intervals. adaSPOCC does well and, importantly, the paper reports the worst-case discrepancy to the best method, which is a nice robustness metric.

Now the soft spots. The three properties are stated as theorems, but the proofs are sketches. Property (a) is fine: a random guess classifier gives uniform conditional columns in the limit, DPT maps those to the constant-one distribution, which is the neutral element. Property (b) as written is not correct. The rectified classifier c_rec = h ∘ c_k, with h(y)=argmax p(Y|C_k=y), is many-to-one in general for multiclass. The paper claims that when nval→∞, the y-th column of M_k equals the y'-th column of M_rec for each x with c_k(x)=y and c_rec(x)=y'. That's only true when h is a bijection; in general the y'-th column of M_rec pools all preimages under h, so the column distributions differ. The stress-test example with three classes and a uniform c_k is concrete and works. So property (b) is not established. Property (c) also has gaps. The proof asserts that HAC will detect exact copies as 'maximally dependent' and use an idempotent t-norm, but the dependence measure κ converges to 1-exp(-H) where H is the entropy of the marginal prediction distribution, which is not 1, and the grid-search heuristic is not proven to select λ=∞ (i.e., the minimum t-norm). The paper also has a typo in that section (T1 instead of T∞), which is minor by itself but not reassuring.

These are addressable. One can probably prove property (b) under an additional condition (h bijective, or a rectification map that preserves the relevant columns), or prove a weaker ensemble invariance. But as published, the central theoretical claim is not backed by the proof. The empirical evidence is good enough to make the method worth knowing and using; just not enough to justify the assertion that it 'possesses' these robustness properties.

I'd recommend engaging with the paper as a serious referee: it deserves a careful review rather than a desk reject. The experiments and the construction are worth publishing after the proofs are fixed.

Referee Report

3 major / 4 minor

Summary. The paper introduces SPOCC, a method for combining the label predictions of an ensemble of classifiers using possibility distributions estimated from validation-set confusion matrices and aggregated with Aczel-Alsina t-norms, and adaSPOCC, an adaptive version that tunes t-norm parameters by a hierarchical clustering heuristic and discounts classifiers by estimated error rate. The central theoretical claim is that, asymptotically as the validation set grows, the aggregation satisfies three robustness properties: (a) random-guess classifiers have no influence, (b) an adversarial (worse-than-random) classifier can be replaced by a rectified sub-random classifier without changing the ensemble output, and (c) exact copies of a classifier have no influence. The paper also presents experiments on synthetic data and eight real datasets comparing against selection, weighted/softmax voting, naive Bayes, Bayes, and stacking.

Significance. If the robustness properties were established, the paper would make a solid contribution: it provides an agnostic, label-based fusion rule with O(ell^2 K) parameter complexity, incremental updates, and formal guarantees that are unusual in this literature. The experimental work is careful and reproducible, with bootstrap confidence intervals, multiple baselines, and both synthetic and real data, and adaSPOCC frequently appears at or near the top. The central theoretical section, however, is not sound as written: the proof of property (b) uses a false column-identity assertion, and the proof of property (c) relies on an unproved selection of lambda = infinity. Because these properties are the paper's main advertised contribution, the current version cannot be recommended for publication.

major comments (3)

[Section 3.7, property (b)] The proof asserts that as nval -> infinity, if ck(x) = y and c_rec(x) = y', then the y-th column of M(k) is identical to the y'-th column of the confusion matrix of c_rec. This is false for many-to-one rectification maps h: the y'-th column of M(rec) is the sum over all j with h(j) = y' of the j-th column of M(k), so it equals M(k)(:,y) only when h^{-1}(y') = {y}. For ell >= 3, many-to-one maps are generic; for example, with conditionals p(Y|C=omega1) = (0.2, 0.3, 0.5), p(Y|C=omega2) = (0.25, 0.25, 0.5), p(Y|C=omega3) = (0.3, 0.3, 0.4), h maps every label to omega3 and the column of M(rec) pools all validation points while the omega1 column of M(k) does not. Therefore the normalized columns, and hence the DPT possibility distributions pi_{k|y} and pi_{rec|y'}, generally differ, and the claimed equality f(c) = f(tilde c) does not follow from the given argument.
[Section 3.7, property (b), and Eq. (13)] Even if the undiscounted possibility distributions of ck and c_rec coincided, adaSPOCC's discounting breaks the claimed invariance. Since r[ck] > (ell-1)/ell and the proof establishes r[c_rec] < 1/ell, Eq. (13) assigns different discounting coefficients alpha_k to the adversarial classifier and to its rectified version. The discounted possibility distributions, which are what the ensemble actually uses, would then differ, so f(c) = f(tilde c) cannot hold for adaSPOCC unless the proof explains why the discounting is immaterial. The current sketch does not address this.
[Section 3.7, property (c), and Appendix A] The proof of property (c) assumes that exact copies are detected as maximally dependent by HAC and are then aggregated with T_infinity = T_min. Two gaps remain. First, the heuristic search in Appendix A selects lambda by validation accuracy, and no argument shows that it will return lambda = infinity for a cluster of exact copies; with any finite lambda, T_lambda(a,a) = a^{2^{1/lambda}} != a, so the two copies do not reduce to a single influence. Second, the constant-classifier case is misstated: if ck is constant, the conditional distribution p(Y | Ck = omega_j) is the marginal p(Y), which is not generally a Dirac distribution, so the claim that copies of a constant classifier yield a Dirac possibility distribution is incorrect.

minor comments (4)

[Section 3.7, property (c)] The text says copies are 'aggregated using T1 = T_min', but Eq. (8) defines T1 = T_times and T_infinity = T_min; this should be corrected.
[Section 2.4, property (b)] The statement 'if ck has an error rate larger than random guess, i.e. r[f(c)] > (ell-1)/ell' should read r[ck] > (ell-1)/ell; the proof uses r[ck].
[Sections 4.2.4, 4.3, and 5] Several typos write 'SPOOC' or 'adaSPOOC' instead of 'SPOCC' or 'adaSPOCC'.
[Table 3] A few entries are missing the 'std.' prefix before standard deviation values (e.g., SPOCC on Drive and adaSPOCC on Avila).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the robustness properties are derived from the DPT/t-norm construction and validation-set estimates, not from fitted constants or self-citations; the proof gaps in Section 3.7 are correctness issues, not circular reductions.

full rationale

The derivation chain is self-contained. Conditional probability estimates are computed from validation-set confusion matrices (Eq. 1), converted via the Dubois-Prade transform (Eq. 5), and combined with Aczel-Alsina t-norms (Eq. 8) whose parameters are tuned on the validation set. The robustness properties (a)-(c) are asymptotic statements about this construction; none of them is defined in terms of a fitted value, and none is obtained by renaming an estimated parameter as a prediction. The paper contains no load-bearing self-citations, no imported uniqueness theorem, and no ansatz smuggled in via prior work by the same authors. The skeptical concerns about Section 3.7 are not circularity: property (b) relies on a column-identity assertion for the rectified classifier's confusion matrix that is false in general for many-to-one rectification maps, and property (c) depends on the grid-search heuristic selecting an idempotent t-norm and contains the inconsistent statement 'T1 = T∧' (elsewhere the paper defines T1 = T× and T∞ = T∧). These are potential gaps in the proofs of the claimed guarantees, but they do not reduce the conclusion to an input by construction or fit. Under the quoted-evidence standard required for flagging circularity, no circular step can be exhibited.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central derivation rests on standard probability, possibility, and t-norm background; the only fitted quantities are the hyperparameters lambda and rho, which adapt the aggregation to dependency and accuracy but do not enter the asymptotic proofs as fitted constants.

free parameters (3)

lambda_a = data-dependent, chosen by grid search on validation set
T-norm strictness hyperparameters for each non-leaf node in the HAC dendrogram; controls how much dependent classifiers are down-weighted (Eq. 8, Appendix A).
rho = data-dependent, chosen by grid search on validation set
Exponent in discounting formula (Eq. 13) that maps validation error rates to discounting coefficients alpha_k.
HAC linkage/threshold = unspecified
The dissimilarity matrix D is clustered with HAC, but the linkage criterion and cophenetic threshold are not fully specified; they affect the tree G and therefore the aggregation.

assumptions (5)

domain assumption The union of the validation fractions used by the meta-learner is an i.i.d. sample of p(X,Y).
Stated in Section 1; needed for MLE consistency of multinomial probability estimates.
standard math Maximum likelihood estimates from the confusion matrix converge almost surely to the true multinomial probabilities as nval tends to infinity.
Used in Section 3.7 to claim asymptotic properties; follows from the strong law of large numbers.
standard math The Dubois-Prade transform (Eq. 5) yields the unique possibility distribution satisfying consistency, preference preservation, and maximal specificity.
Invoked in Section 3.2 and Appendix B.2, relying on results from [12].
standard math Aczel-Alsina t-norms T_lambda form a family with T_1 = product and T_infinity = minimum, and have 1 as neutral element.
Used in Sections 3.3 and 3.4; standard t-norm theory.
domain assumption Base classifiers are agnostic and only output a class label; score-based aggregation is out of scope.
Stated in Section 2.3; defines the setup but limits applicability.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SPOCC: Scalable POssibilistic Classifier Combination -- toward robust aggregation of classifiers." pith.science (2026). https://pith.science/paper/46WIZKR3

@misc{pith2026190806475,
  author       = {Pith},
  title        = {Pith review of: SPOCC: Scalable POssibilistic Classifier Combination -- toward robust aggregation of classifiers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/46WIZKR3}},
  note         = {Machine review of arXiv:1908.06475}
}

read the original abstract

We investigate a problem in which each member of a group of learners is trained separately to solve the same classification task. Each learner has access to a training dataset (possibly with overlap across learners) but each trained classifier can be evaluated on a validation dataset. We propose a new approach to aggregate the learner predictions in the possibility theory framework. For each classifier prediction, we build a possibility distribution assessing how likely the classifier prediction is correct using frequentist probabilities estimated on the validation set. The possibility distributions are aggregated using an adaptive t-norm that can accommodate dependency and poor accuracy of the classifier predictions. We prove that the proposed approach possesses a number of desirable classifier combination robustness properties.

Figures

Figures reproduced from arXiv: 1908.06475 by the authors.

**Figure 1.** Example of a dendrogram for K = 5. Leaf nodes are at the bottom. For each of the four non-leaf nodes, a specific dependence level λa (a ∈ {1; 2; 3; 4}) must be determined to compute the aggregated possibility distribution. HAC relies on a classifier dissimilarity matrix D. In our case, entries of this K × K matrix are simply given by Dkk0 = 1 − κ (ck, ck0 ). The t-norm based possibility distribution aggregation met… view at source ↗

**Figure 2.** Synthetic dataset obtained from four Gaussian distributions. Exam [PITH_FULL_IMAGE:figures/full_fig_p018_2.png] view at source ↗

**Figure 3.** Subsets of the data seen respectively by [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗

Figures from the paper (7 more)

**Figure 4.** Figure 4: Evolution of accuracy distributions (violin plots) for several aggrega [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]

**Figure 5.** Figure 5: Evolution of accuracy distributions (violin plots) for several aggrega [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]

**Figure 6.** Figure 6: Evolution of accuracy distributions (violin plots) for several aggrega [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]

**Figure 7.** Figure 7: Normalized averaged confusion matrices for the reported results in [PITH_FULL_IMAGE:figures/full_fig_p023_7.png]

**Figure 8.** Figure 8: Evolution of accuracy distributions (violin plots) on an heterogeneous [PITH_FULL_IMAGE:figures/full_fig_p024_8.png]

**Figure 9.** Figure 9: Evolution of accuracy distributions (violin plots) with class label [PITH_FULL_IMAGE:figures/full_fig_p025_9.png]

**Figure 10.** Figure 10: Evolution of possibility values of πens w.r.t. to λ The above example illustrates that depending on the value of λ, the class label achieving maximal possibility is either a class label strongly supported by one classifier or moderately supported by a majority of clas…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 51 canonical work pages

[1]

Alsinet, C

T. Alsinet, C. I. Ches˜ nevar, L. Godo, and G. R. Simari. A logic program- ming framework for possibilistic argumentation: Formalization and logical properties. Fuzzy Sets and Systems , 159(10):1208–1228, 2008. 57

work page 2008
[2]

N. B. Amor, D. Dubois, H. Gouider, and H. Prade. Possibilistic preference networks. Information Sciences, 460:401–415, 2018

work page 2018
[3]

Balakrishnan and M

N. Balakrishnan and M. Mojirsheibani. A simple method for combining estimates to improve the overall error rates in classiﬁcation. Computational Statistics, 30(4):1033–1049, 2015

work page 2015
[4]

Bella, C

A. Bella, C. Ferri, J. Hern´ andez-Orallo, and M. J. Ram´ ırez-Quintana. On the eﬀect of calibration in classiﬁer combination. Applied intelligence , 38(4):566–585, 2013

work page 2013
[5]

Benferhat, D

S. Benferhat, D. Dubois, L. Garcia, and H. Prade. Possibilistic logic bases and possibilistic graphs. In Proceedings of the Fifteenth conference on Un- certainty in artiﬁcial intelligence, pages 57–64. Morgan Kaufmann Publish- ers Inc., 1999

work page 1999
[6]

G. Biau, A. Fischer, B. Guedj, and J. D. Malley. Cobra: A combined regression strategy. Journal of Multivariate Analysis , 146:18–28, 2016

work page 2016
[7]

L. Breiman. Bagging predictors. Machine learning, 24(2):123–140, 1996

work page 1996
[8]

Brown, J

G. Brown, J. Wyatt, R. Harris, and X. Yao. Diversity creation methods: a survey and categorisation. Information Fusion, 6(1):5–20, 2005

work page 2005

Show all 52 references

[9]

De Cooman and D

G. De Cooman and D. Aeyels. Supremum preserving upper probabilities. Information Sciences, 118(1-4):173–212, 1999

1999
[10]

Deerwester, S

S. Deerwester, S. T. Dumais, G. W. Furnas, T. K. Landauer, and R. Harsh- man. Indexing by latent semantic analysis. Journal of the American society for information science , 41(6):391–407, 1990

1990
[11]

Destercke, D

S. Destercke, D. Dubois, and E. Chojnacki. Possibilistic information fusion using maximal coherent subsets. IEEE Transactions on Fuzzy Systems , 17(1):79–92, 2008

2008
[12]

Dubois, L

D. Dubois, L. Foulloy, G. Mauris, and H. Prade. Probability-possibility transformations, triangular fuzzy sets, and probabilistic inequalities. Reli- able computing, 10(4):273–297, 2004

2004
[13]

Dubois, J

D. Dubois, J. Lang, and H. Prade. Automated reasoning using possibilistic logic: Semantics, belief revision, and variable certainty weights. IEEE transactions on knowledge and data engineering , 6(1):64–71, 1994

1994
[14]

Dubois and H

D. Dubois and H. Prade. On several representations of an uncertain body of evidence. Fuzzy Information and Decision Processes, pages 161–181, 1982

1982
[15]

Dubois and H

D. Dubois and H. Prade. Possibility theory: An approach to the computer- ized processing of information. Plenum Press, New York, 1988

1988
[16]

Dubois and H

D. Dubois and H. Prade. When upper probabilities are possibility measures. Fuzzy sets and systems , 49(1):65–74, 1992

1992
[17]

Dubois and H

D. Dubois and H. Prade. Possibility theory and its applications: Where do we stand? In Springer Handbook of Computational Intelligence , pages 31–60. Springer, 2015. 58

2015
[18]

Dubois, H

D. Dubois, H. Prade, and S. Schockaert. Generalized possibilistic logic: foundations and applications to qualitative reasoning about uncertainty. Artiﬁcial Intelligence, 252:139–174, 2017

2017
[19]

I. Gilboa. Expected utility with purely subjective non-additive probabili- ties. Journal of Mathematical Economics , 16(1):65 – 88, 1987

1987
[20]

I. R. Goodman. Fuzzy sets as equivalence classes of random sets. In R. Yager, editor, Fuzzy Sets and Possibility Theory , pages pp. 327–343. Pergamon Press, Oxford, 1982

1982
[21]

K. Guo, T. Xu, X. Kui, R. Zhang, and T. Chi. ifusion: Towards eﬃcient intelligence fusion for deep learning from real-time and heterogeneous data. Information Fusion, 51:215–223, 2019

2019
[22]

Hoang, N

M. Hoang, N. Hoang, B. K. H. Low, and C. Kingsford. Collective model fusion for multiple black-box experts. In International Conference on Ma- chine Learning, pages 2742–2750, 2019

2019
[23]

Y. S. Huang and C. Y. Suen. A method of combining multiple experts for the recognition of unconstrained handwritten numerals. IEEE transactions on pattern analysis and machine intelligence , 17(1):90–94, 1995

1995
[24]

H¨ ullermeier

E. H¨ ullermeier. Possibilistic induction in decision-tree learning. InEuropean Conference on Machine Learning, pages 173–184. Springer, 2002

2002
[25]

J. Ji, X. Chen, Q. Wang, L. Yu, and P. Li. Learning to learn gradient aggregation by gradient descent. In Proceedings of the Twenty-Eighth In- ternational Joint Conference on Artiﬁcial Intelligence, IJCAI-19 , pages 2614–2620. International Joint Conferences on Artiﬁcial Intel...

2019
[26]

Jousselme and P

A.-L. Jousselme and P. Maupin. Distances in evidence theory: Compre- hensive survey and generalizations. International Journal of Approximate Reasoning, 53(2):118 – 145, 2012. Theory of Belief Functions (BELIEF 2010)

2012
[27]

Kamp´ e de F´ eriet

J. Kamp´ e de F´ eriet. Interpretation of membership functions of fuzzy sets in terms of plausibility and belief. In E. S. Madan M. Gupta, editor, Fuzzy Information and Decision Processes , pages 93–98. North-Holland, Amsterdam, 1982

1982
[28]

Kim and Z

H.-C. Kim and Z. Ghahramani. Bayesian classiﬁer combination. In Artiﬁ- cial Intelligence and Statistics , pages 619–627, 2012

2012
[29]

Kim and J

T. Kim and J. Ghosh. On single source robustness in deep fusion models. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d’ Alch´ e-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 4815–4826. Curran Associates, Inc., 2019

2019
[30]

Lacoste, M

A. Lacoste, M. Marchand, F. Laviolette, and H. Larochelle. Agnostic bayesian learning of ensembles. In International Conference on Machine Learning, pages 611–619, 2014. 59

2014
[31]

L. Li, W. Xu, T. Chen, G. B. Giannakis, and Q. Ling. Rsa: Byzantine- robust stochastic aggregation methods for distributed learning from het- erogeneous datasets. In Proceedings of the AAAI Conference on Artiﬁcial Intelligence, volume 33, pages 1544–1551, 2019

2019
[32]

Y. Li, B. Rubinstein, and T. Cohn. Exploiting worker correlation for label aggregation in crowdsourcing. In International Conference on Machine Learning, pages 3886–3895, 2019

2019
[33]

Liu and L

H. Liu and L. Zhang. Advancing ensemble learning performance through data transformation and classiﬁers fusion in granular computing context. Expert Systems with Applications , 131:20–29, 2019

2019
[34]

S. Loustau. Aggregation of svm classiﬁers using sobolev spaces. Journal of Machine Learning Research, 9(Jul):1559–1582, 2008

2008
[35]

X. Ma, C. Ji, X. Zhang, J. Wang, J. Li, and K.-C. Li. Secure multiparty learning from aggregation of locally trained models. In International Con- ference on Machine Learning for Cyber Security , pages 173–182. Springer, 2019

2019
[36]

Madry, A

A. Madry, A. Makelov, L. Schmidt, D. Tsipras, and A. Vladu. Towards deep learning models resistant to adversarial attacks. In International Con- ference on Learning Representations, 2018

2018
[37]

A. K. Menon, A. Rajagopalan, B. Sumengen, G. Citovsky, Q. Cao, and S. Kumar. Online hierarchical clustering approximations. arXiv preprint arXiv:1909.09667, 2019

1909 arXiv
[38]

Pei-Zhuang and E

W. Pei-Zhuang and E. Sanchez. Treating a fuzzy subset as a projectable random subset. In M. Gupta and E. Sanchez, editors, Fuzzy Information and Decision Processes, pages 213–220. North-Holland, Amsterdam, 1982

1982
[39]

D. D.-H. Prade. Possibilistic logic, preferential models, non-monotonicity and related issues. In Proc. of IJCAI, volume 91, pages 419–424, 1991

1991
[40]

Rigollet and A

P. Rigollet and A. B. Tsybakov. Sparse estimation by exponential weight- ing. Statistical Science, pages 558–575, 2012

2012
[41]

T. Sales. Fuzzy sets as set classes. Stochastica, 6(3):249–264, 1982

1982
[42]

L. J. Savage. The foundations of statistics. NY, John Wiley, pages 188–190, 1954

1954
[43]

R. E. Schapire. The strength of weak learnability. Machine learning , 5(2):197–227, 1990

1990
[44]

Serrurier and H

M. Serrurier and H. Prade. Entropy evaluation based on conﬁdence inter- vals of frequency estimates: Application to the learning of decision trees. In International Conference on Machine Learning , pages 1576–1584, 2015

2015
[45]

G. Shafer. A Mathematical Theory of Evidence. Princeton University press, Princeton (NJ), USA, 1976. 60

1976
[46]

P. P. Shenoy. Using possibility theory in expert systems. Fuzzy Sets and Systems, 52(2):129–142, 1992

1992
[47]

Tulyakov, S

S. Tulyakov, S. Jaeger, V. Govindaraju, and D. Doermann. Review of classiﬁer combination methods. In Machine learning in document analysis and recognition, pages 361–386. Springer, 2008

2008
[48]

J. H. Ward Jr. Hierarchical grouping to optimize an objective function. Journal of the American statistical association , 58(301):236–244, 1963

1963
[49]

D. H. Wolpert. Stacked generalization. Neural networks , 5(2):241–259, 1992

1992
[50]

Wozniak, M

M. Wozniak, M. Grana, and E. Corchado. A survey of multiple classiﬁer systems as hybrid systems. Information Fusion, 16:3 – 17, 2014. Special Issue on Information Fusion in Hybrid Intelligent Fusion Systems

2014
[51]

X. Yao, T. Huang, R.-X. Zhang, R. Li, and L. Sun. Federated learning with unbiased gradient aggregation and controllable meta updating. In NeuRIPS Workshop on Federated Learning, page to appear, 2019

2019
[52]

L. A. Zadeh. Fuzzy sets as a basis for a theory of possibility. Fuzzy Sets and Systems, 1:3–28, 1978. 61

1978

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Alsinet, C

T. Alsinet, C. I. Ches˜ nevar, L. Godo, and G. R. Simari. A logic program- ming framework for possibilistic argumentation: Formalization and logical properties. Fuzzy Sets and Systems , 159(10):1208–1228, 2008. 57

work page 2008

[2] [2]

N. B. Amor, D. Dubois, H. Gouider, and H. Prade. Possibilistic preference networks. Information Sciences, 460:401–415, 2018

work page 2018

[3] [3]

Balakrishnan and M

N. Balakrishnan and M. Mojirsheibani. A simple method for combining estimates to improve the overall error rates in classiﬁcation. Computational Statistics, 30(4):1033–1049, 2015

work page 2015

[4] [4]

Bella, C

A. Bella, C. Ferri, J. Hern´ andez-Orallo, and M. J. Ram´ ırez-Quintana. On the eﬀect of calibration in classiﬁer combination. Applied intelligence , 38(4):566–585, 2013

work page 2013

[5] [5]

Benferhat, D

S. Benferhat, D. Dubois, L. Garcia, and H. Prade. Possibilistic logic bases and possibilistic graphs. In Proceedings of the Fifteenth conference on Un- certainty in artiﬁcial intelligence, pages 57–64. Morgan Kaufmann Publish- ers Inc., 1999

work page 1999

[6] [6]

G. Biau, A. Fischer, B. Guedj, and J. D. Malley. Cobra: A combined regression strategy. Journal of Multivariate Analysis , 146:18–28, 2016

work page 2016

[7] [7]

L. Breiman. Bagging predictors. Machine learning, 24(2):123–140, 1996

work page 1996

[8] [8]

Brown, J

G. Brown, J. Wyatt, R. Harris, and X. Yao. Diversity creation methods: a survey and categorisation. Information Fusion, 6(1):5–20, 2005

work page 2005