acceptodds
Under review as a conference paper at ICLR 2027

Concept-Aware View Routing with Sparse Compensation for Multi-View Multi-Label Feature Selection

Abstract

In recent years, multi-view multi-label learning (MVML) has attracted increasing attention due to its ability to characterize complex real-world objects by exploiting complementary information from multiple views and rich multi-label semantics. However, selecting discriminative features from high-dimensional multi-view data remains a critical challenge in MVML. Most existing methods adopt unified view fusion and shared feature selection strategies, making it difficult to capture the heterogeneous view dependencies associated with different label semantics and potentially overlooking specific features required for challenging semantic concepts. To address these issues, we propose View-Conditioned Route Selection for Multi-View Multi-Label Feature Selection (VCRS). Specifically, VCRS first constructs a semantic space that integrates atomic concepts and relational concepts, thereby jointly preserving label information and higher-order label semantics. Concept confidence, graph constraints, and adaptive view routing are further incorporated to achieve fine-grained concept-level multi-view fusion. In addition, VCRS learns a shared set of base features via nonconvex sparse learning, identifies under-represented concepts according to their concept reconstruction errors and importance, and selects concept-specific features through a sparse compensation mechanism. In this way, VCRS enhances the feature representation of challenging semantic concepts while retaining the advantages of shared feature selection. Experimental results on multiple multi-view datasets demonstrate the effectiveness of the proposed method.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-view multi-label learning, View-conditioned routing, Concept-specific feature selection

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_WSs3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

VCRS selects features for multi-view multi-label data. It combines original labels with clustered label-group responses, learns concept-specific view weights, and then applies sparse corrections to concepts that are poorly represented by the base features. Experiments on six datasets report improvements over six baselines.

Critical strengths
  1. The motivation is reasonable. Shared feature selection can miss information needed by particular labels, and different labels may depend on different views.
  2. The method achieves the highest reported AP on all six datasets. The ablations provide initial evidence that the proposed components help.
Critical weaknesses
  1. The proposed mechanisms lack direct evidence. Overall gains do not show that concepts use more suitable views or that compensation helps poorly represented concepts. Relational concepts are averages of grouped labels, and the ablation removes several components together. Without separate tests or concept-level analysis, it is unclear whether the method works for the reasons claimed.
  2. The motivation is already present in work on view-specific labels and label-specific features. The paper does not clearly establish what its routing and compensation add, and comparisons with the closest methods are missing. This makes the contribution difficult to judge.
Questions for authors
  1. What evidence shows that routing selects suitable views and that compensation improves the identified concepts on held-out data? Do separate ablations or a comparison with random label groups support the proposed explanation?
  2. What is the specific advance over existing methods using view-specific labels or label-specific features, and what evidence supports that distinction?
1 comment

We thank the reviewer for recognizing our research motivation and experimental performance. We provide the following clarifications and additional information in response to the comments. To examine whether routing selects more suitable views, we added an equal-cardinality view-replacement experiment in Appendix F to complement the ablation study in the main text. Rather than evaluating overall predictive performance, we summarize concept reconstruction errors on an independent test set. On yeast, VOC07, and MIRFlickr, the aggregated concept reconstruction errors of VCRS are 0.8018, 0.9286, and 0.7979, respectively, all lower than the 0.8313, 0.9391, and 0.8533 obtained with alternative routing, representing an average relative reduction of 3.72%. Thus, the improvement is associated with concept-conditioned view selection rather than simply combining more views. Consistently, Table 4 shows that removing the routing module reduces AP on MIRFlickr from 0.7760 to 0.7648.

To assess the effectiveness of the compensation targets, Appendix G.1 examines the relationship between concept reconstruction error and label prediction difficulty on an independent test set. The Spearman correlation coefficients on yeast, VOC07, and MIRFlickr are 0.9508, 0.9038, and 0.9663, respectively, indicating that concept reconstruction error is an effective audit criterion for identifying harder-to-predict concepts. To distinguish targeted compensation from the benefits of compensation in general, we also conducted an equal-size random-compensation control. We matched the numbers of compensated atomic and relational concepts separately to those in VCRS, while randomly selecting compensation targets within each category. To control for additional Stage II iterations, we also evaluated no_comp. For each seed, it shares VCRS's Stage I solution and continues updating HH, BB, AA, and WW for the same number of Stage II iterations, while fixing R(v)=0R^{(v)}=0 for all views. The AP results (mean ± std) on four datasets with different label-space sizes are as follows:

Datasetno_compRandomVCRS
yeast0.7587±0.0080.7587±0.0080.7604±0.008
VOC070.6990±0.0200.7001±0.0190.7011±0.019
MIRFlickr0.7753±0.0150.7752±0.0150.7760±0.011
Corel5K0.3823±0.0360.3922±0.0410.3974±0.045

VCRS achieves higher AP than both no_comp and Random on all four datasets. In particular, VCRS's sparse compensation mechanism achieves more pronounced performance gains on Corel5K, which has a larger label space and provides a more discriminative evaluation setting. Together, these findings support the value of concept auditing and sparse compensation.

To clarify the distinction from methods using view-specific labels or label-specific features, we take DHLI, as discussed in the main text, as an example. Whereas DHLI partitions labels into common, view-specific, and noisy categories, VCRS operates at the granularity of atomic and relational concepts. It adaptively allocates view information using observability constraints OvkO_{vk} and continuous routing weights AvkA_{vk}, and, after shared row-sparse base feature selection, activates sparse compensation RR only for audited under-represented concepts, rather than assigning labels to fixed views or applying compensation indiscriminately. This also distinguishes VCRS from approaches that only learn label-specific features. Under the same experimental protocol, DHLI and VCRS were each tuned using validation AP. The mean test AP results (DHLI / VCRS) on emotions, yeast, Corel5K, SCENE, VOC07, and MIRFlickr, respectively, are 0.7161 / 0.7871, 0.7432 / 0.7604, 0.3682 / 0.3974, 0.8493 / 0.8571, 0.6639 / 0.7011, and 0.7622 / 0.7760. VCRS achieves higher AP on all six datasets, with an average improvement of 2.94 percentage points.

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

This paper proposes VCRS, a two-stage framework for multi-view multi-label feature selection that integrates concept-aware semantic modeling, adaptive view routing, and sparse compensation. The paper provides a relatively complete objective formulation and optimization procedure. However, several key components rely on heuristic designs and lack sufficient theoretical justification and experimental validation.

Critical strengths

1.The overall optimization framework is relatively complete. Appendix D provides detailed alternating update rules, making the optimization procedure reasonably clear. 2.The paper identifies a meaningful limitation of global view fusion, where different semantic concepts may depend on different views. The two-stage design provides a reasonable approach to balancing shared feature selection and concept-specific requirements.

Critical weaknesses
  1. Limited novelty of higher-order label interaction modeling. Although the paper emphasizes higher-order label semantics, the relational concepts are constructed through linear averaging of label vectors. The proof in Appendix A establishes that these concepts do not introduce additional independent dimensions into the linear representation space. However, it remains unclear whether the proposed method captures higher-order label interactions beyond group-level aggregation of existing label dependencies. Moreover, advantages over conventional label grouping and structured semantic modeling also remain insufficiently established.
  2. The novelty of relational concepts and sparse compensation is not sufficiently established. Relational concepts are essentially constructed through label clustering and within-cluster averaging. While it does not establish new statistical interactions beyond the original label space. Furthermore, the compensation branch in Eq. (8) essentially introduces additive linear corrections rather than a new expressive representation.
  3. The ablation results provide limited evidence for some key components. In Table 3, the AP improvements of the full VCRS over VCRSv3 are relatively small. On MIRFlickr, removing the concept modeling components or sparse compensation leads to only marginal performance degradation. These results raise questions about the individual contributions and necessity of the proposed components, particularly sparse compensation.
  4. Table 3 only reports overall AP and HL, whereas the main motivation of sparse compensation is to improve the representation of under-represented concepts. The current experiments do not directly demonstrate whether these concepts actually benefit from compensation.
Questions for authors
  1. Could the authors verify whether the concept reconstruction error ek effectively identifies concepts with poor predictive performance? 2.Please provide a comparison with a simpler single-stage structured sparse learning model to demonstrate the necessity of the proposed two-stage compensation strategy.
  2. Please provide potential view-dimensionality bias, and concept confidence.
  3. Please clarify the specific advantages of the proposed routing and compensation strategy over other concept-conditioned sparse models with comparable complexity.
1 comment

We thank the reviewer for recognizing the significance of our research question and optimization framework. In response to the comments, we provide the following clarifications and additional information. To assess the effectiveness of concept auditing, Appendix G.1 examines the relationship between concept reconstruction error and label prediction difficulty on an independent test set. On yeast, VOC07, and MIRFlickr, the Spearman rank correlation coefficients are 0.9508, 0.9038, and 0.9663, respectively. The results show that concepts with larger reconstruction errors are generally harder to predict, supporting the use of eke_k to identify concepts that are difficult to predict using shared features. The concept audit further combines concept frequency and semantic coverage into qkq_k to determine which concepts receive incremental compensation.

To examine the necessity of the two-stage strategy, we added a supplementary experiment in Appendix G.2. The single-stage variant retains the other VCRS mechanisms but jointly optimizes WW and RR from the first iteration, without performing an inter-stage concept audit; its total number of outer iterations matches that of two-stage VCRS. The results are summarized below (VCRS / single-stage VCRS):

DatasetAP ↑HL ↓Running time (s) ↓Speedup
yeast0.7604 / 0.75870.2043 / 0.205315.53 / 17.021.10×
SCENE0.8571 / 0.85640.0877 / 0.088356.71 / 653.1211.52×
Corel5K0.3974 / 0.37330.0124 / 0.01271,066.45 / 20,541.6719.26×

These results indicate that the gains stem not only from additional compensation but also from the two-stage optimization strategy, which first learns stable shared base features and then selectively activates compensation based on concept auditing. Across datasets of different scales, the single-stage variant does not improve performance over two-stage VCRS and, particularly on larger datasets, incurs several-fold higher running time, further supporting the necessity of the two-stage sparse compensation strategy.

Regarding potential view-dimensionality bias, we use the absolute Pearson correlation between individual features and concepts to mitigate differences in feature-value scales. For each concept, the observability threshold is defined relative to its strongest view; view contributions are then learned through simplex-constrained AvkA_{vk} rather than weighted directly by view dimensionality. Concept confidence is not assigned uniformly by hand, either. Atomic concepts correspond directly to ground-truth labels and have confidence 11; for relational concepts, we take the equally weighted average of within-group label-response agreement and the stability of member-label positive rates, then clip it to [0.25,0.90][0.25, 0.90]. This is a heuristic weight based on response consistency.

Unlike methods that learn only common/view-specific labels or use a unified sparse weighting scheme, VCRS combines concept-level adaptive view routing with selective sparse compensation. The former allocates suitable view information to different concepts, while the latter adds necessary features only for under-represented concepts, avoiding indiscriminate increases in model complexity. Beyond the ablation studies in the main text, Appendix F evaluates different view subsets while keeping the number of candidate views per concept fixed, whereas Appendix G.2 compares two-stage VCRS with a single-stage variant using the same optimization variables. Together, these results support the effectiveness of this joint design of concept-level adaptive view routing and selective sparse compensation.

Finally, we have clarified the scope and substance of our methodological contribution in the main text. Relational concepts do not increase the rank of the label matrix, nor do we claim that they uncover irreducible three-way or higher-order interactions. Instead, they provide structured coordinates for group-level relationships beyond atomic concepts, allowing these relationships to be jointly supervised, routed, and audited.

Reviewer_MZq7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes VCRS for multi-view multi-label feature selection. It builds a concept space containing the original labels and label-group representations, learns concept-dependent view weights, and selects shared base features through row-sparse regularization. A second stage identifies poorly represented concepts and introduces sparse compensation. Experiments evaluate the method on six datasets using four multi-label metrics and feature budgets ranging from 1% to 20%.

Critical strengths
  1. The method addresses a sensible limitation of shared feature selection. Different labels can depend on different views, and features useful for less common concepts may receive insufficient weight in a shared selector. VCRS gives these ideas a concrete formulation through concept-dependent routing and a separate sparse compensation branch. The two-stage procedure also provides a clear way to identify compensation targets from the base solution.
  2. The overall experimental evidence is favorable. VCRS achieves the highest mean AP on all six datasets and the best numerical result in most dataset–metric combinations. The gains are most visible on Corel5K and VOC07, while the feature-budget curves show that the results are not confined to a single subset size. The use of a common downstream classifier, repeated runs, and training-only preprocessing makes the comparisons reasonably informative.
Critical weaknesses
  1. My main concern is that the experiments do not establish the specific contribution of targeted compensation. The audit in Section 3.5 selects concepts using reconstruction error and frequency, but Table 3 reports only aggregate performance. It does not show whether the selected concepts benefit, or whether the audit outperforms an equally sized random compensation set. Moreover, Eq. (8) adds the compensation coefficients to the base coefficients, and both continue to be optimized in Stage II, leaving the source of the improvement unclear. The reported AP gains from compensation are 0.17–0.26 percentage points, without uncertainty estimates for the ablation differences. Given the importance of this mechanism to the paper’s contribution, stronger controlled evidence is needed to justify the design and the claim that its components are indispensable.
  2. The parameter-selection procedure needs fuller reporting. Equation (10) contains several coupled objectives controlled by four main weights. Section 4.1 gives a search range, and Figure 3 varies each parameter while holding the others fixed on MIRFlickr. However, the manuscript does not specify the validation criterion, joint search procedure, or final settings for each dataset. Appendix E establishes a training/validation/test separation, but does not resolve these choices. These details matter for reproducing the comparisons and understanding the tuning effort required.
Questions for authors
  1. Do the audited concepts benefit more from compensation than the remaining concepts? Can the authors compare the proposed audit with an equally sized random compensation set and report variability across runs?
  2. How were the weights in Eq. (10) selected for each dataset? Please specify the validation criterion and search procedure, and clarify whether the loss terms were normalized to account for differences in scale.
  3. How does concept-conditioned routing differ from the view-specific label modeling used by DHLI, already discussed in Section 2.2? A clearer explanation of this distinction would help position the contribution; a matched comparison would also be useful.
2 comments

We sincerely thank the reviewer for the valuable suggestions. We have carefully considered the comments and provide the following clarifications and additional information. To examine whether concept auditing is more effective than arbitrary assignment of compensation targets, we conducted an equal-size Random control experiment. For each run, we randomly selected compensation targets separately from the atomic and relational concept sets, matching the number selected by VCRS in each set, and repeated this random target-set selection three times. To control for additional Stage II iterations, we also evaluated no_comp. For each seed, it shares VCRS's Stage I solution and continues updating HH, BB, AA, and WW for the same number of Stage II iterations, while fixing R(v)=0R^{(v)}=0 for all views. The AP results (mean ± std) on four datasets with different label-space sizes are as follows:

Datasetno_compRandomVCRS
yeast0.7587±0.0080.7587±0.0080.7604±0.008
VOC070.6990±0.0200.7001±0.0190.7011±0.019
MIRFlickr0.7753±0.0150.7752±0.0150.7760±0.011
Corel5K0.3823±0.0360.3922±0.0410.3974±0.045

VCRS achieves higher mean AP than both no_comp and Random on all four datasets, providing further evidence that audit-selected compensation targets yield additional gains over equal-size random assignment. Further analysis of the experimental logs reveals substantial overlap between randomly selected targets and those identified by the VCRS concept audit on datasets with relatively few labels, with a mean target overlap of 51.3%. This helps explain the modest gains over Random on these datasets, whereas a more pronounced improvement emerges on Corel5K, which has a larger label space. We also added a supplementary experiment in Appendix G.1 to examine the relationship between concept reconstruction error and label prediction difficulty on an independent test set. The Spearman rank correlation coefficients on yeast, VOC07, and MIRFlickr are 0.9508, 0.9038, and 0.9663, respectively, confirming the effectiveness of the audit criterion.

For the four weights in Eq. (10), we use a single-pass coordinate-wise grid search rather than exhaustive four-dimensional joint search. Following the order α→β→γ→λ\alpha\to\beta\to\gamma\to\lambda, each weight is selected from {10−3,10−2,…,103}\{10^{-3},10^{-2},\ldots,10^3\} to maximize mean validation AP across five random seeds and feature ratios from 1% to 20%. The final (α,β,γ,λ)(\alpha,\beta,\gamma,\lambda) values recorded in the experimental logs are: emotions (0.1, 1, 1, 1), yeast (0.01, 1, 1, 10), Corel5K (0.1, 10, 1, 0.001), SCENE (0.1, 1, 1, 0.1), VOC07 (1, 1, 1, 0.001), and MIRFlickr (0.1, 10, 1, 0.001). Input features undergo MinMax normalization. The loss and regularization terms in Eq. (10) are computed as written, without additional term-wise scale normalization; the four weights calibrate their relative contributions through the above validation procedure.

Moreover, the key distinction between VCRS and DHLI lies not in whether they account for view-dependent label heterogeneity, but in the granularity of modeling and allocation. DHLI categorizes labels as common, view-specific, or noisy, using the view-specific labels to guide feature selection within each view. In contrast, VCRS retains the original labels as atomic concepts and introduces relational concepts. For each concept, observability OvkO_{vk} restricts the eligible views, continuous routing weights AvkA_{vk} are learned under simplex constraints, and sparse compensation is activated only for audited under-represented concepts. Under the same experimental protocol, both methods are tuned independently using validation AP. The mean test AP values for DHLI / VCRS on emotions, yeast, Corel5K, SCENE, VOC07, and MIRFlickr are 0.7161 / 0.7871, 0.7432 / 0.7604, 0.3682 / 0.3974, 0.8493 / 0.8571, 0.6639 / 0.7011, and 0.7622 / 0.7760, respectively. VCRS achieves higher AP on all six datasets, with an average improvement of 2.94 percentage points.

Thank you for your response. The additional controls, tuning details, and DHLI comparison address several of my concerns. The response strengthens the empirical evidence, but the remaining uncertainty about the added value of audit-guided compensation leaves my overall assessment unchanged. I therefore maintain my original rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.