acceptodds
Under review as a conference paper at ICLR 2027

Continuous-Aware Multi-view Attribute Weighted Naive Bayes

Abstract

Attribute weighting provides a flexible way to alleviate the conditional independence assumption in naive Bayes. Within this widely studied area, multi-view attribute weighting has emerged as a new paradigm and achieved state-of-the-art performance by exploiting complementary information from multiple data representations. However, existing multi-view attribute weighting methods are primarily designed for nominal attributes, whereas continuous attributes are more prevalent in real-world applications. Although discretization enables these methods to handle continuous attributes, it removes differences between values within the same interval, limiting their ability to capture fine-grained continuous information. To address this issue, we propose a novel method called Continuous-aware Multi-view Attribute Weighted Naive Bayes (CMAWNB), which directly models continuous attributes by constructing three additional views from the perspective of complex attribute dependencies, continuous space partitions, and local group structures, respectively. Subsequently, attribute weights are independently optimized within each view, and the view-specific predictions are fused for classification. Experiments on 55 benchmark continuous datasets demonstrate the effectiveness of CMAWNB.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
multi-view learning, continuous attributes, attribute weighting, naive Bayes

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ushHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces CMAWNB, a multi-view extension of weighted naive Bayes aimed at continuous attributes. The starting observation is that existing multi-view weighted NB variants typically discretize the inputs before building their auxiliary label views, which throws away the differences between values that happen to fall in the same bin. CMAWNB keeps the original continuous attributes as a raw view and constructs three categorical label views from the predictions of Gaussian super-parent one-dependence estimators, random trees grown with continuous splits, and granular-ball k-NN classifiers. The raw view is modeled with class-specific attribute-weighted Gaussian NB, while each label view uses class-specific attribute-value weighted NB; the weights are fit by minimizing conditional log-likelihood, with an L2 penalty pulling the label-view weights toward one. The four views are combined by averaging their posterior probabilities. The method is evaluated on 55 UCI datasets with ten independent runs of ten-fold cross-validation. CMAWNB obtains the best average rank among the eight compared methods and statistically significant aggregate gains over all of them. The component-wise and incremental ablations point to the continuous random trees and the granular-ball view as the main sources of improvement, and a synthetic case study illustrates how the granular balls help track nonlinear local class structure.

Critical strengths
  1. The problem being addressed is real, and the remedy is sensible. Attribute weighting cannot recover information that discretization has already discarded, so carrying continuous values through both the raw model and the label-view construction is well motivated. The three label views also encode genuinely different inductive biases—class-conditional linear dependence, axis-aligned partitioning, and distance-based local structure—which is what a multi-view construction should do.

  2. The empirical evaluation is unusually broad for this line of work: 55 datasets, dataset-level tables, average ranks, win/loss counts, and Wilcoxon signed-rank tests. The component-wise and incremental ablations give a reasonable picture of where the gains come from, and the aggregate margin over PMWNB, the closest predecessor, is encouraging.

Critical weaknesses
  1. Limited methodological novelty. CMAWNB is essentially a careful assembly of components that already exist in this literature: Gaussian weighted NB, GSPODE, random trees, granular-ball k-NN, attribute-value weighting, and plain probability averaging. The advance over MAWNB-style predecessors is a continuous-data adaptation plus integration, not a new learning mechanism. That is not disqualifying by itself, but the incremental nature of the contribution should be acknowledged more plainly.

  2. Little analysis beyond the optimization details. The gradient derivations show how the weights are found, but nothing is said about statistical consistency, generalization, calibration of the averaged posteriors, when the views are actually complementary, or under what conditions fusion should beat the best single view. Restricting all weights to [0,1] and regularizing label-view weights toward one are likewise adopted without justification or sensitivity analysis.

  3. The training protocol for label views is underspecified, and as described it risks being an improper form of stacking. Nowhere does the paper say whether the GSPODE, random-tree, and GBKNN predictions used to fit the second-level CAVWNB models are generated out of fold. If a view constructor is fit before the outer split, that is straightforward leakage. Even if the constructors are fit only on the outer-training fold, fitting the second-level model on in-sample predictions and evaluating it on out-of-sample predictions creates a representation-distribution mismatch. This worry is largest for the random trees and the label-dependent granular-ball partitions, both of which can fit training points closely. If an inner out-of-fold protocol was in fact used, it needs to be stated; if not, the reported gains may be optimistic.

  4. Dataset and preprocessing descriptions are unclear and in places internally inconsistent. The paper states that all 55 datasets contain only continuous attributes, yet monks-1 is standardly a categorical dataset. The australian entry is reported with 12,546 instances, eight attributes, and three classes, which does not correspond to the commonly used Australian credit dataset. The two energy-efficiency targets are turned into 37- and 38-class problems with no explanation of how the labels were constructed. The case study also calls fourclass a "four-class dataset" while stating it contains two classes. As written, the benchmark cannot be independently reconstructed.

  5. The evaluation stays almost entirely within the weighted-NB family. Besides random forest, no general-purpose continuous-data classifier is included—no gradient-boosted trees, RBF-SVM, k-NN, or any tabular neural model. Given that the method trains m GSPODEs, m random trees, and m granular-ball partitions before fitting four weighted classifiers, the complete absence of cost analysis (training time, inference time, memory, scaling with m and n) makes it difficult to judge whether the added complexity buys enough accuracy to be worth it.

Questions for authors
  1. How are the label attributes for training instances produced within each outer cross-validation fold? Are they out-of-fold predictions from an inner loop, or predictions from base learners trained on the very instances they predict? It would also help to confirm that imputation, normalization, discretization, tree growing, and granular-ball construction are all fit after the outer split, never before it.

  2. The dataset descriptions need to be pinned down before the benchmark can be reproduced. For australian, monks-1, both energy-efficiency targets, and the fourclass case study, could you provide dataset identifiers or download links, the exact target-construction steps, and the preprocessing applied? Several settings are also missing from the paper: the value and selection procedure for λ, probability smoothing, variance floors for the Gaussian estimates, random-tree stopping criteria, and any safeguards governing granular-ball termination.

  3. Would the main findings survive two changes to the experimental design: (a) regenerating every training label view through inner out-of-fold prediction, and (b) adding strong continuous-data baselines such as tuned gradient-boosted trees, RBF-SVM, k-NN, and a modern tabular neural model? Since the method trains 3m base learners plus four weighted classifiers, reporting training time, inference time, and memory alongside accuracy would make the accuracy–cost trade-off much easier to assess.

3 comments
AuthorsOpenReview ↗

Response to Reviewer <ushH> (1/2)

W1. Methodological novelty. CMAWNB preserves numerical distinctions that input discretization would discard and constructs three complementary label views to capture structures not explicitly modeled by the raw view. The GSPODE, random tree, and granular ball views respectively capture class conditional attribute dependencies, threshold based partitions, and local class structure. To mitigate attribute redundancy, CMAWNB learns attribute weights within each view. Together, these design choices form a multi-view weighted naive Bayes framework for continuous data.

W2. View complementarity and weight regularization. The ablation shows that mean accuracy increases at each step. In the instance-level error correction analysis, restoring each omitted view yields a positive mean net accuracy gain. Please see lines 454–472 of our revised paper for the new instance-level error correction analysis. Across the 55 datasets, fusion outperforms every standalone view in mean accuracy. These results support view complementarity.

MethodRawGSPODERTGBCMAWNB
Mean accuracy (%)77.7178.7983.0178.7184.14

We constrain weights to [0, 1] so that each log-likelihood term can be downweighted without being amplified 1 leaves it unchanged and 0 removes it. Because attribute-value weighting introduces more parameters in the label views, we center the L2 penalty at 1 to discourage large deviations from the unweighted model.

W3 & Q3(a). In-sample versus out-of-fold label generation. In the original CMAWNB, the base classifiers were fitted only on the outer training fold and used to generate label attributes for those same training instances. This can create a distribution mismatch between the training and test label attributes. To assess this effect, we evaluated CMAWNB-OOF using inner 10-fold cross-validation. Each training instance’s label attributes were generated by base classifiers that had not seen it, while the outer test procedure and equal averaging remained unchanged.

DatasetOriginal CMAWNB (%)CMAWNB-OOF (%)
artificial-characters94.3794.16
climate90.6791.06
energy-efficiency-y164.9266.82
wine-white69.2868.60

OOF generation increased mean accuracy by 0.35 percentage points across the four datasets. Thus, replacing in-sample labels with OOF labels did not consistently reduce test accuracy, but required ten additional fits of each label view constructor per outer fold.

W4. Dataset identificationWe treated the integer-coded categorical predictors in MONK’s-1 as numerical inputs. The entry australian refers to Australian Sign Language, not Australian Credit (https://raw.githubusercontent.com/icesky0125/dataset-and-raw-results-for-HGS-paper/master/datasets/sign.arff). For Energy Efficiency, we rounded both targets to the nearest integer: for Y1Y_1, ties were rounded to even and the 37 distinct values were mapped to class IDs 1–37 in lexicographic order; for Y2Y_2, ties were rounded up and the integers 11–48 were used directly as class labels. Finally, fourclass is the dataset name, not the class count: the LIBSVM version we used has two classes, 862 instances, and two attributes.

AuthorsOpenReview ↗

Response to Reviewer <ushH> (2/2)

W5 & Q3(b). Computational cost and additional baselines. Let nn, mm, and KK denote the numbers of training instances, attributes, and classes, respectively. We separate training into label-view construction (GG) and view-specific weight learning (TT).

  • Label-view construction. Fitting the base classifiers and generating the three label representations for all training instances costs G=O(mn(D+1)[q(n+K)+m+K]+HIKnm2+n[Km2+nm2+(m+K)M]).G = O\left(mn(D+1)[q(n+K)+m+K] + HIKnm^2 + n[Km^2+nm^2+(m+K)M]\right). Here, DD, qq, HH, and II are the observed data-dependent maxima for tree depth, attributes examined per node, splitting rounds per partition, and Lloyd iterations per split, respectively; MM is the total number of final granular balls.

  • Weight learning. Probability estimation and weight optimization across the four views cost T=O((1+L)(nmK+mK2))T=O((1+L)(nmK+mK^2)), where LL is the total number of objective and gradient evaluations.

The overall training cost is G+TG+T. CMAWNB incurs additional runtime but achieved higher mean classification accuracy in our benchmark comparisons. We consider this trade-off acceptable when predictive accuracy is the priority.

We added XGBoost and RBF-SVM baselines using the same 10 runs of 10-fold cross-validation on the 55 datasets. CMAWNB achieved comparable mean accuracy to XGBoost and higher mean accuracy than the evaluated RBF-SVM.

MethodCMAWNBXGBoostRBF-SVM
Mean accuracy (%)84.1484.0764.34

Q1. Preprocessing and outer-fold fitting. Before cross-validation, we computed imputation means on each full dataset and determined equal-width bin boundaries only for baselines requiring discretization. After each outer split, we grew the random trees, constructed the granular ball partitions, and computed the attribute ranges for GBKNN distance normalization using only the training fold. The same ranges were used at test time.

Q2. Implementation settings and stopping criteria. We use the default L2L_2 regularization coefficient λ=1\lambda=1 for label-view weighting, without tuning. We apply Laplace smoothing with a pseudocount of 1. The Gaussian variance floor is (0.01/6)2(0.01/6)^2. A random tree node stops splitting if it is pure, its total instance weight is below 2, or no candidate split yields information gain greater than 10−610^{-6}. Granular balls with purity of at least 0.9 are not split. Only nonempty child balls are retained, and splitting stops when a full round does not increase the number of balls.

Thank you for the additional experiments and clarifications. The new analyses improve the paper, particularly regarding view complementarity. However, concerns about in-sample label generation, fold-wise preprocessing, and the limited evaluation of additional baselines remain. The methodological novelty is also still relatively incremental. Overall, I appreciate the authors' efforts, but these changes are not sufficient to alter my original assessment. I maintain my rating.

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

This paper addresses the problem that existing multi-view attribute-weighted naive Bayes methods rely on discretization to handle continuous data, thereby losing fine-grained numerical information within the same interval. The authors propose CMAWNB, which retains the original continuous view and constructs three complementary label views based on GSPODE, random trees, and granular ball KNN, modeling class-conditional linear attribute dependencies, continuous space threshold partitions, and local instance group structures, respectively. The raw view uses class-specific attribute weighting, while the three label views use class-specific attribute-value weighting with L2 regularization. The four views are fused by averaging class-membership probabilities. Experiments on 55 continuous-attribute UCI datasets include comparison, ablation, parameter sensitivity, and visualization case studies, validating the effectiveness of each component and reporting performance advantages over existing multi-view naive Bayes models.

Critical strengths

A targeted multi-view modeling solution to the information loss of continuous attributes.

Critical weaknesses

-Conventional weight learning and fusion strategies.

-Lack of justification for view selection and complementarity analysis.

-Lack of computational complexity and runtime analysis.

Questions for authors

-In Section 3.2, for each training sample, are base classifiers trained on a subset that excludes that sample? If not, how is overfitting of base classifiers to training samples prevented from affecting second-stage attribute-weight learning?

-Please justify the choice of the three base classifiers and consider analyzing error correlation, Q-statistic, disagreement measure, Kappa, and the relationship between individual view performance and fused performance. In addition, have other continuous-data classifiers been tried for constructing label views?If replacing one base classifier still yields similar gains, the advantage may stem mainly from multi-model ensembling rather than the specific view design?

-Please provide time and space complexity analysis, training time, inference time, and memory usage curves under different sample sizes and attribute dimensionalities, and runtime comparisons with MAWNB, PMWNB, and ICGNB under the same experimental environment and implementation platform.

-Whether the imputation means, discretization boundaries, and any data-dependent model parameters are computed solely from the current training fold without using test-fold information. In addition, please provide sensitivity analyses for the granular-ball purity threshold (rho=0.9), regularization parameter (lambda), number of candidate attributes in random trees, number of k-means initializations and random seeds in GBKNN, and the number of base classifiers.

3 comments
AuthorsOpenReview ↗

Response to Reviewer <adti> (1/2)

W1. Weight learning and fusion strategy. CMAWNB’s main contribution is the construction of complementary views that capture different characteristics of the same training instances. We use established methods to learn attribute weights within each view. We also evaluated adaptive fusion, which weights each view according to the probabilities it assigns to the true labels of nearby training instances.

DatasetEqual averaging (%)Adaptive weighting (%)
artificial-characters94.3794.72
climate90.6790.59
iris96.0096.07
wine-white69.2868.72

The two fusion strategies achieved comparable accuracy across the four datasets. We therefore retained equal averaging as a simpler, competitive default that requires no additional view-weight estimation.

W2. Ablation and view complementarity. The incremental ablation shows an accuracy gain at each step. To further assess view complementarity, we added a subsection on instance-level error correction, where we compare CMAWNB with variants that omit one view at prediction time . Across the 55 datasets, restoring the Raw, GSPODE, RT, and GB views yields mean net accuracy gains of 0.38, 0.79, 1.78, and 0.17 percentage points, respectively. Please see lines 454–472 of our revised paper for the new instance-level error correction analysis. These results support the complementary role of each view in CMAWNB’s fusion.

W3 & Q3. Computational complexity and runtime.

Time complexity. Let nn, mm, and KK denote the numbers of training instances, attributes, and classes, respectively. We separate training into label-view construction (GG) and view-specific weight learning (TT).

  • Label-view construction. Fitting the base classifiers and generating the three label representations for all training instances costs G=O(mn(D+1)[q(n+K)+m+K]+HIKnm2+n[Km2+nm2+(m+K)M]).G = O\left(mn(D+1)[q(n+K)+m+K] + HIKnm^2 + n[Km^2+nm^2+(m+K)M]\right). Here, DD, qq, HH, and II are the observed data-dependent maxima for tree depth, attributes examined per node, splitting rounds per partition, and Lloyd iterations per split, respectively; MM is the total number of final granular balls.
  • Weight learning. Probability estimation and weight optimization across the four views cost T=O((1+L)(nmK+mK2))T=O((1+L)(nmK+mK^2)), where LL is the total number of objective and gradient evaluations.

The overall training cost is G+TG+T. CMAWNB incurs additional runtime but achieved higher mean classification accuracy in our benchmark comparisons. We consider this trade-off acceptable when predictive accuracy is the priority.

Space complexity. Let SS be the total number of nodes across all random trees. The fitted model uses O(nm+Km2+mK2+KS)O(nm+Km^2+mK^2+KS) storage for the retained label view data and ball memberships, GSPODE statistics, weighted NB tables, and random trees. Weight optimization additionally uses up to O(nmK)O(nmK) working space.

Measured runtime. We compared CMAWNB and PMWNB on six datasets that vary in the number of instances and attributes. The table reports mean per-fold wall-clock times over ten runs of ten-fold cross-validation. Values displayed as 0.00 s are rounded, not necessarily zero.

GroupDatasetInstancesAttributesCMAWNB train (s)PMWNB train (s)CMAWNB test (s)PMWNB test (s)
Number of instancesfertility10090.010.000.000.00
Number of instanceswine-red1,599110.761.460.060.01
Number of instancesartificial-characters10,21879.6117.370.810.44
Number of attributesuser40350.020.080.000.00
Number of attributesspectf-heart349440.280.170.020.00
Number of attributeslibras360902.708.230.200.14
AuthorsOpenReview ↗

Response to Reviewer <adti> (2/2)

Q1. In-sample versus out-of-fold label generation. In the original CMAWNB, the base classifiers were fitted on the outer training fold and used to generate label attributes for those same instances. To assess this risk, we evaluated CMAWNB-OOF. Within each outer training fold, we used inner 10-fold cross-validation so that each instance’s label attributes were generated by models that had not seen it. The test-time procedure remained unchanged. This required ten additional fits per label-view constructor.

DatasetOriginal CMAWNB (%)CMAWNB-OOF (%)
artificial-characters94.3794.16
climate90.6791.06
energy-efficiency-y164.9266.82
wine-white69.2868.60

OOF generation increased training cost and yielded limited, inconsistent accuracy gains across the four datasets.

Q2. Base classifier selection and view complementarity. We chose GSPODEs, RTs, and GBKNNs to capture distinct structures in numerical data: class-conditional attribute dependencies, threshold based partitions, and local class structure, respectively.

Across the 55 benchmark datasets, mean disagreement rates ranged from 15.33% to 20.29% across the three pairs of label views. This diversity in error patterns, together with the positive mean net accuracy gain from restoring each omitted view, supports the complementary role of the views in CMAWNB.

View pairMean disagreement rate (%)
GSPODE–RT16.48
GSPODE–GB20.29
RT–GB15.33

Across the 55 datasets, fusing all four views in CMAWNB yielded higher mean accuracy than using any single view alone.

MethodRawGSPODERTGBCMAWNB
Mean accuracy (%)77.7178.7983.0178.7184.14

Q4. Preprocessing protocol and hyperparameter sensitivity. Before cross-validation, we computed imputation means on each full dataset and, for methods requiring discretization, determined equal-width bin boundaries on the same full dataset. All model parameters were then fitted using only the corresponding training fold.

We varied ρ\rho while keeping all other settings fixed. The table reports mean classification accuracy (%). Across the tested thresholds, accuracy varied by only 0.15 percentage points on climate and 0.73 percentage points on iris, indicating low sensitivity to ρ\rho on these two datasets.

Datasetρ=0.20\rho=0.20ρ=0.40\rho=0.40ρ=0.60\rho=0.60ρ=0.80\rho=0.80ρ=0.85\rho=0.85ρ=0.90\rho=0.90ρ=0.95\rho=0.95ρ=1.00\rho=1.00
climate90.6790.6790.6790.6790.6790.6790.6390.52
iris96.0795.8095.8096.2095.9396.0095.6795.47

We adopt the regularization scheme of the existing attribute-value weighting method with λ=1\lambda=1. For random trees, the number of candidate attributes considered at each split follows prior work. Since GSPODE uses one base classifier per input attribute, we use mm base classifiers in each of the RT and GBKNN views to keep all three label views the same width. Please see lines 161–163 of our revised paper for the rationale behind using mm base classifiers in each label view. In GBKNN, different kk-means initializations produce separate partitions, one per base classifier. The random seed is fixed for reproducibility and is not tuned for accuracy.

Reviewer_adtiOpenReview ↗

Official Comment by Reviewer adti

Thank the authors for the responses. I will keep my rating.

Reviewer_DB3ZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Continuous-Aware Multi-view Attribute Weighted Naive Bayes (CMAWNB) to address the information loss caused by discretizing continuous attributes in existing multi-view NB methods. It preserves the raw continuous view and constructs three complementary label views using GSPODEs, random trees, and GBKNNs to capture attribute dependencies, space partitions, and local class structures. Experiments on 55 benchmark datasets, along with ablation studies and a case study, evaluate the proposed method's effectiveness.

Critical strengths
  1. The paper has a reasonable motivation. It aims to preserve fine-grained information that may be lost when continuous attributes are discretized in Naive Bayes methods.
  2. The design of the different views is interesting, including local class information. The case study also provides visual evidence of its effectiveness.
  3. The experiments cover 55 benchmark datasets, providing a broad evaluation.
Critical weaknesses
  1. The novelty of the proposed method could be further clarified, particularly regarding the specific contribution of the granular-ball view compared with existing neighborhood-based approaches.
  2. The four-view design requires further justification, including the necessity and complementarity of each view, as well as the use of equal-weight fusion.
  3. Although the experiments support the effectiveness of continuous modeling, more direct comparisons with discretized modeling would help strengthen this claim. The selection of baselines and discretization settings also requires further justification.
Questions for authors
  1. The rationale behind some parameter choices in the method is unclear. For example, why is the number of base classifiers for RT and GBKNN set to (m)?
  2. According to Eq. (16), the four views are fused by simply averaging their predicted probabilities. Could differences in predictive performance or redundancy among views affect the final results? If so, can the current weighting mechanism handle these issues?
  3. The competing methods that require nominal inputs use equal-width discretization with ten bins. Is this choice supported by prior studies? Have other bin sizes or discretization strategies been evaluated to ensure a fair comparison?
  4. The paper lists several baselines but does not categorize them, making it difficult to assess the coverage of the comparison. Could the authors clarify the selection criteria and explain why more recent related methods, if any, and strong classifiers such as XGBoost were not included?
  5. The progressive ablation only considers one order of combining the components. Could the authors also evaluate other view combinations or leave-one-view-out variants to justify the four-view design better?
2 comments
AuthorsOpenReview ↗

Response to Reviewer <DB3Z>

W1. Comparison with k-NN. Standard KNN generates a label by voting over the class labels of the k nearest training instances to each query. In contrast, CMAWNB forms local groups during training using a class-purity criterion and, in each partition, uses the majority-class label of the granular ball closest to the query as a label attribute. The innovation lies in turning group-level class information from multiple partitions into a label view with learned class-specific attribute-value weights. Please see lines 66–69 of our revised paper for the distinction between standard k-NN and the granular-ball label view.

W2. Four-view design and complementarity. The component-wise ablation shows that each proposed modification improves mean accuracy over MAWNB, while the incremental ablation shows further gains at each step. To assess complementarity directly, we added a paired instance-level analysis in which we omit one view at prediction time without retraining. Across 55 datasets, restoring the Raw, GSPODE, RT, and GB views yields mean net accuracy gains of 0.38, 0.79, 1.78, and 0.17 percentage points, respectively. Please see lines 454–472 of our revised paper for the new instance-level error correction analysis.

W3. Continuous versus discretized modeling. We compared CMAWNB with MAWNB, PMWNB, and CAVWNB, all of which use discretization in our evaluation. To isolate the effect of continuous modeling, we used discretized inputs for the raw, SPODE, and random-tree branches while keeping the granular ball view and equal weight fusion unchanged. Across 55 datasets, this variant achieved 82.25% mean accuracy, compared with 84.14% for CMAWNB.

Q1. Number of base classifiers. GSPODE uses one super-parent per input attribute, yielding m base classifiers. We likewise use m random trees and m GBKNN classifiers to give all three label views the same width without introducing a separate hyperparameter. Please see lines 161–163 of our revised paper for the rationale behind using mm base classifiers in each label view.

Q2 & W2. Equal weight versus adaptive fusion. Equal averaging combines complementary predictions without additional view-level parameters, providing a simple, robustness-oriented fusion rule. We also tested EMAWNB-style adaptive fusion, which estimates view weights from distance-weighted true-class probabilities on the 30 nearest training instances.

DatasetEqual averaging (%)Adaptive weighting (%)
artificial-characters94.3794.72
climate90.6790.59
iris96.0096.07
wine-white69.2868.72

Under the same 10 runs of 10-fold cross-validation, adaptive weighting showed no consistent improvement, supporting equal averaging as a simple, competitive default.

Q3. Choice of discretization settings. Following Wu et al. (2026, Sec. 4.1), we applied 10-bin equal-width discretization to all baselines requiring nominal inputs. This is the default setting of WEKA’s unsupervised Discretize filter. Equal-width discretization is a common preprocessing method for continuous attributes.

Q4. Baseline selection and coverage. We have revised the experimental section to group the seven baselines by their roles: PMWNB and MAWNB are the closest multi view attribute weighted NB methods; CAVWNB and WANBIA represent single view attribute weighting; ICGNB and GNB directly model continuous attributes; and RF provides a tree-based baseline. The original comparison already includes two recent related methods, PMWNB (2026) and ICGNB (2025). Please see lines 359–365 of our revised paper for the revised categorization of the seven original baselines. To broaden the comparison beyond these methods, we have now also evaluated XGBoost and RBF-SVM on the same 55 datasets using the same 10-run, 10-fold cross validation protocol.

MethodCMAWNBXGBoostRBF-SVM
Mean accuracy (%)84.1484.0764.34

Q5. Leave-one-view-out ablation. To address the order dependence of the progressive ablation, we conducted a leave-one-view-out experiment on “wine-white”. Starting from the CMAWNB model, we removed each view separately, keeping the other views and the equal-averaging rule unchanged.

VariantMean accuracy (%)
CMAWNB69.28
Without the raw view68.79
Without the GSPODE view68.71
Without the random-tree view65.54
Without the granular-ball view67.53
Removing any view reduced accuracy on “wine-white”, with the largest drop after removing the random-tree view.

Thank the authors for the responses and the additional experimental results. The added analyses address some of my concerns about the view design and continuous modeling. However, considering my concerns about the distinct contribution of the granular-ball view and the fairness of the experimental comparisons, I will keep my rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.