- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This article investigates the feature selection problem for incomplete multi-view partial-label data, where each sample's candidate label set contains exactly one real label. Faced with the problem that traditional single view biased methods may not be effective in multi-view scenario. The article proposes the IMPLFS method, which jointly learns view-specific similarity graphs, cross-view graph reconstruction coefficients, an instance-adaptive fused graph, label confidences constrained to the candidate sets, and sparse feature-selection matrices. In the feature distance and label reconstruction term, missing observation data is masked, while the cross view graph reconstruction mechanism achieves the transfer of structural information between views. The objective function is optimized by alternately solving the simplex constrained quadratic programming problem and iteratively reweighting the feature update process. The experiment was conducted on four benchmark datasets, combined with four downstream label classifiers, to examine the feature selection performance under different view loss rates and feature retention ratios. The results show that among all 16 combinations of classifiers and datasets, this method achieved the highest average accuracy; The ablation experiment also validated the effectiveness of complementary reconstruction and label disambiguation mechanisms.
Critical strengths
1.The modules are logically interconnected, and the overall framework is coherent.
This paper constructs a unified optimization framework to address three interrelated challenges: missing views, candidate label ambiguity, and feature redundancy. Cross-view graph reconstruction leverages relational information from other views to refine intra-view structures; the fused graph provides neighborhood information for label confidence estimation, which in turn supervises sparse feature learning. Clear information flow exists between these modules, enabling structural learning, label disambiguation, and feature selection to work in concert rather than being handled in isolation. From a modeling perspective, there is a clear correspondence between the individual objective terms and the specific problems being addressed. 2.The experimental analysis covers a wide range of dimensions, examining the effectiveness of the method from multiple perspectives.
The paper not only compares the classification performance of various feature selection methods but also evaluates the suitability of the selected features using four downstream partial-label classifiers, thereby reducing reliance on any single specific classifier. Furthermore, the experiments encompass a range of missingness rates, feature retention ratios, parameter settings, and component ablations. As shown in Table 2, the proposed method achieves the highest average accuracy across all 16 combinations of classifiers and datasets; ablation studies also demonstrate that removing either the complementary reconstruction component or the label disambiguation component leads to a decline in performance. These results provide consistent empirical support for the method's effectiveness and the roles of its two primary components.
Critical weaknesses
1.The substantive methodological innovations have not been sufficiently distinguished or substantiated.
While the research problem holds practical significance, the proposed method primarily integrates cross-view graph reconstruction, adaptive graph fusion, graph-smoothing label disambiguation, and row-sparse feature learning into a unified objective. The main text fails to adequately articulate the specific technical contributions—whether in modeling or optimization—that the proposed framework offers compared to existing applications of these mechanisms.
In particular, both cross-view reconstruction and adaptive fusion operate by constraining inter-graph relationships; although they serve distinct functions, their roles may overlap. While the ablation studies validate the benefits of complementary reconstruction and label disambiguation, they do not directly demonstrate the necessity of sample-level adaptive fusion, nor do they sufficiently distinguish the effects of joint optimization from the simple concatenation of existing modules. Consequently, the current evidence supports the effectiveness of the combination, yet the justification for its unique methodological contributions remains insufficient.
2.There is insufficient analysis and validation regarding graph learning biases and label disambiguation errors caused by missing data and unreliable structural information.
In Equation (1), the feature distance is weighted by the view availability indicator variable . Consequently, if either endpoint is missing in the current view, the distance cost for that sample pair is set to zero. For an observed sample, connections to observed endpoints with positive feature distances incur a cost, whereas connections to missing endpoints do not.
This treatment may make edges involving missing endpoints more likely to be assigned similarity weights, thereby influencing subsequent label disambiguation. Although the model incorporates label information via the constraint , the differentiation between candidate labels relies primarily on graph smoothness and feature fitting. If the handling of missing data creates erroneous neighborhoods—or if noisy features and structural conflicts across views cause neighborhood relationships to span across true class boundaries—label confidence may skew toward incorrect candidates. Since label confidence also influences the integrated graph learning and feature selection processes, these errors can be further amplified during joint optimization.
While cross-view reconstruction, sparse regularization, and other objective terms might mitigate these issues, the paper fails to adequately explain how these mechanisms suppress the connection bias associated with missing endpoints, nor does it provide diagnostics of the graph structure or experiments involving structural perturbations. Consequently, the existing experiments involving random data loss are insufficient to demonstrate whether the method can maintain graph quality and prevent erroneous disambiguation—and its subsequent propagation—when structural information is unreliable.
Questions for authors
1.On Innovation and Module Necessity.
Please clearly articulate the novel technical contributions of this paper compared to existing mechanisms for graph reconstruction, graph fusion, and sparse feature learning. Under identical experimental conditions, does joint optimization yield additional benefits over a staged processing approach? What degree of improvement does sample-level adaptive fusion offer compared to uniform fusion or global view-weighting fusion?
2.On the Sufficiency of Comparisons.
Could you include comparisons with suitably adapted recent methods for partial-label feature selection and incomplete multi-view feature selection, ensuring consistent missingness settings, feature budgets, and comparable parameter choices? Additionally, how do baseline approaches—such as using the full feature set or performing feature selection after missing-view imputation—perform? These comparisons would help assess the actual benefits of the proposed joint mechanism.
3.On Handling Missing Data and Structural Reliability
In Equation (1), setting the distance cost associated with missing endpoints to zero might assign them disproportionately high similarity weights, potentially affecting label disambiguation; how do the cross-view reconstruction and other regularization terms mitigate this bias and the resulting erroneous feedback? Could you combine connection weight statistics with controlled structural perturbation experiments to report metrics on graph quality, label disambiguation quality, and classification performance, thereby evaluating the method's robustness against missing endpoints, noisy views, or erroneous neighborhood structures?
Response to Reviewer PLbv on Joint Optimization, Baseline Comparisons, and HOG Perturbation
Thank you for reading our manuscript carefully and identifying which conclusions needed further experimental support. Your suggestions helped us strengthen the evaluation and clarify what the results say about our design.
Q1. Our contribution is a joint feature-selection objective for incomplete multi-view partial-label data. Cross-view reconstruction recovers view-specific sample relations, and the fused graph guides candidate-label disambiguation. The resulting label confidences then inform the graph updates and supervise sparse feature selection.
Following your suggestion, we added two comparisons in Section 4.7 (Table 5), both at 30% missingness and a 40% feature budget. The first compares full IMPLFS with staged optimization. Staged recovers and fuses the view graphs first, then performs label disambiguation and feature selection with the fused graph fixed. The second compares sample-specific, uniform, and global fusion. These schemes learn weights for each sample, assign equal weights to the views, and share one set of weights across all samples, respectively.
The results show that compared to the staged baseline, complete IMPLFS learns substantially cleaner fused graphs across all datasets: without joint supervisory feedback from label confidences, the staged variant retains 86.53%--93.87% off-diagonal cross-class edge weight in L, whereas joint IMPLFS reduces it to 21.38%--74.87% (a reduction of over 65 percentage points on both Yahoo!News100 and Yahoo!News50) while achieving higher classification accuracy on most datasets. Comparing the fusion strategies , sample-level fusion outperforms uniform fusion on all four datasets, and performs comparably to global fusion overall with slight gains on most datasets. Together with the ablation and extended comparison results in the revised manuscript (Tables 3–5), these findings confirm that the performance gains stem primarily from end-to-end joint coupling rather than trivial stacking of individual components.
Q2. We evaluated IMPLFS against recent feature selectors, using all features, and selecting features after imputation.
Following your suggestion, we added TIME-FS, TRUST-FS, and TERUIMUFS to Table 2, along with PLFS, a partial-label feature-selection control obtained by removing missing-view reconstruction and multi-view processing from IMPLFS. The methods use matched data splits, missingness masks, and feature budgets. The three incomplete multi-view baselines use their authors' example settings. Section 4.6 (Table 3) adds the all-feature and two-stage comparisons. In the two-stage comparisons, we impute the missing views using either cross-view k-NN or graphs learned by the IMPLFS graph module, then use PLFS for feature selection.
Across the expanded comparisons in Table 2, IMPLFS achieves the highest mean accuracy in the vast majority of settings. In Table 3, IMPLFS-graph imputation outperforms cross-view k-NN imputation in 7/8 settings; however, even when assisted by the IMPLFS graph module for imputation, the two-stage pipeline still trails complete IMPLFS across all 8/8 settings (by 1.50 to 17.03 percentage points) under the same feature budget. Moreover, using only 40% of the features, complete IMPLFS matches or outperforms using 100% full features in 5/8 settings.
These results further confirm the clear advantage of end-to-end joint optimization over decoupled imputation and feature selection.
Q3. As you noted, zero distance costs for pairs with missing endpoints in Eq. (1) can bias the learned edge weights and affect label disambiguation. The cross-view reconstruction term still constrains these weights through the other view graphs, and the squared Frobenius penalty discourages weight concentration. We checked whether the preference remains and whether it grows when HOG features are shuffled.
Following your suggestion, we added missing-endpoint edge-weight diagnostics and controlled HOG perturbation experiments (Appendix Table F.2). Under the same 30% missingness setting, we randomly permuted the HOG feature rows of 30% of observed training instances and compared original IMPLFS before and after perturbation. Across all four datasets, the edge-weight preference toward missing targets is slightly positive yet negligible in magnitude (Δw ranges from +0.0008 to +0.0042). Within the reported precision of Table F.2, both Δw and training label disambiguation accuracy remain unchanged after perturbation; the off-diagonal cross-class edge weight share of the fused graph shifts by at most 0.02 percentage points, and the mean test accuracies of SAFE-PLL and PL-A-kNN vary by at most 1.13 and 0.79 percentage points, respectively.
These results confirm that the missing-endpoint bias is effectively controlled within a negligible range, and IMPLFS maintains strong stability in graph quality, label disambiguation, and downstream classification under this structural perturbation.