- 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
-
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.
-
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
-
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.
-
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.
-
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.
-
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.
-
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
-
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.
-
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.
-
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
Response to Reviewer <ushH> (2/2)
W5 & Q3(b). Computational cost and additional baselines. Let , , and denote the numbers of training instances, attributes, and classes, respectively. We separate training into label-view construction () and view-specific weight learning ().
-
Label-view construction. Fitting the base classifiers and generating the three label representations for all training instances costs Here, , , , and are the observed data-dependent maxima for tree depth, attributes examined per node, splitting rounds per partition, and Lloyd iterations per split, respectively; is the total number of final granular balls.
-
Weight learning. Probability estimation and weight optimization across the four views cost , where is the total number of objective and gradient evaluations.
The overall training cost is . 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.
| Method | CMAWNB | XGBoost | RBF-SVM |
|---|---|---|---|
| Mean accuracy (%) | 84.14 | 84.07 | 64.34 |
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 regularization coefficient for label-view weighting, without tuning. We apply Laplace smoothing with a pseudocount of 1. The Gaussian variance floor is . 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 . 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.
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.
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.
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 Y1, ties were rounded to even and the 37 distinct values were mapped to class IDs 1–37 in lexicographic order; for Y2, 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.