acceptodds
Withdrawn from ICLR 2027

Simplex-Gated Tensor-Product Regularization for Multi-View Multi-Label Feature Selection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-view multi-label feature selection requires balancing shared label structure with view-specific predictive information. We propose Simplex-Gated TensorProduct Regularization (SG-TPR), which preserves the underlying linear featureto-label function class while organizing shared task structure through tensorproduct role–filler geometry. Each view-specific mapping combines a stabilized data-driven residual component with a simplex-gated low-rank structural component, where feature-side roles interact with shared label-side fillers. Signed training-label correlations initialize and calibrate the shared label geometry, while joint row-energy sparsity yields a unified global feature ranking. Across seven benchmarks and eight competitive baselines evaluated under a unified protocol, SG-TPR achieves the best cross-dataset average rank on Average Precision, Coverage, Hamming Loss, and Ranking Loss, with 21 first-place and three secondplace results among 28 dataset–metric comparisons. Matched component controls and an independently retuned gate control support the structural branch and adaptive simplex allocation. These results are consistent with the proposed tensorproduct-organized structured regularization while preserving the underlying linear function class.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-view Multi-label Feature Selection, Tensor-Product Regularization, Simplex-Gated Structured Learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_AoeDOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SG-TPR for multi-view multi-label feature selection. The method decomposes each view-specific prediction mapping into a residual branch and a shared low-rank structural branch. A simplex gate adjusts the strength of structural regularization across views, while label correlations are used to constrain the shared label factor. The method then combines the two branches to produce a global feature ranking. Experiments on seven datasets compare SG-TPR with eight baselines and include several component ablations.

Critical strengths
  1. The method combines shared label structure with view-specific information while preserving the full class of linear prediction mappings. The theoretical analysis of this property is clear.

  2. The evaluation covers seven datasets and eight baselines under a unified protocol. The paper also includes several ablations to examine the contributions of its main components.

Critical weaknesses
  1. The novelty is limited. The paper argues that existing methods struggle to balance shared label structure and view-specific information, as well as adaptive sharing across views. However, these issues have already been addressed to some extent by methods such as DHLI [1] and SPLIT [2]. Low-rank factorization, view weighting, and sparse feature selection are also well-established techniques. Furthermore, the tensor-product formulation highlighted in the title is simply an algebraic rewriting of a standard inner product, and the method does not introduce an explicit regularizer for tensor structure.

  2. The performance gains are small. Despite sequentially tuning four hyperparameters to select the best configuration, the AP improvement over the strongest baseline is only on the order of 0.001 for most datasets. The statistical tests in Appendix E also fail to establish statistically significant improvements.

  3. The paper does not adequately explain how the proposed components address specific limitations of existing methods. Figure 1 mixes mathematical expressions with ordinary text, contains poorly formatted formulas, and includes several graphical elements whose purpose is unclear. The large box around Eq. (10) is also inconsistent with the formatting of other equations.

[1] Pingting Hao, Kunpeng Liu, and Wanfu Gao. Double-Layer Hybrid-Label Identification Feature Selection for Multi-View Multi-Label Learning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 38(11):12295–12303, 2024.

[2] Lu Sun, Canh Hao Nguyen, and Hiroshi Mamitsuka. Multiplicative Sparse Feature Decomposition for Efficient Multi-View Multi-Task Learning. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI), pages 3506–3512, 2019.

Questions for authors
  1. Most ablation experiments reuse the hyperparameters selected for the full model rather than tuning each variant independently. Even under this setting, removing some components leads to only minor performance differences. Could the authors provide results with independently tuned hyperparameters for each ablation variant to better establish the necessity of these components?

  2. Appendix G.2 shows that the learned view weights π\pi vary considerably across data splits, making their interpretation unclear. Moreover, replacing the learned gate with uniform weights and independently retuning the model reduces average AP by only about 0.005. What meaningful role does the simplex gate actually play?

Reviewer_CwcUOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Simplex-Gated Tensor-Product Regularization (SG-TPR) for multi-view multi-label feature selection. The method combines a view-specific residual branch with a shared low-rank structural branch, while using simplex gating and signed label correlations to adaptively model shared label relationships. Experiments on seven datasets against eight baselines demonstrate competitive performance across four evaluation metrics, with ablation studies supporting the effectiveness of the main components.

Critical strengths
  1. Well-motivated model design with preserved linear expressiveness. The proposed framework combines view-specific residual learning with shared low-rank structural modeling. This design allows the model to exploit shared label relationships while preserving view-specific predictive information. The theoretical demonstration that the full linear function class is preserved further supports the proposed formulation.

  2. Comprehensive experimental evaluation and well-designed ablation studies. The paper evaluates the proposed method against eight baselines on seven datasets using four metrics. The experiments follow a unified evaluation protocol, and several ablation studies investigate the contributions of individual components. The independently retuned Uniform-Gate control further strengthens the experimental evidence.

Critical weaknesses
  1. The novelty of the tensor-product formulation is not sufficiently clear. The proposed tensor-product representation can be directly expressed through conventional bilinear matrix factorization. Therefore, the main contribution appears to be the combination of existing regularization mechanisms rather than a fundamentally new tensor formulation. More direct comparisons with conventional structured multi-task regularization approaches would help clarify the unique benefits of the proposed design.

  2. The scalability and computational efficiency are insufficiently evaluated. The evaluated datasets are relatively small, with at most approximately 6,000 samples and 38 labels. Although theoretical complexity is discussed, direct runtime and memory comparisons are missing. It remains unclear whether the method can efficiently scale to larger and higher-dimensional datasets.

  3. The lack of deep learning comparisons limits the evaluation of nonlinear modeling capabilities. Multi-view multi-label learning often involves complex nonlinear interactions among features, labels, and feature-label relationships. The proposed linear framework may struggle to capture such dependencies compared with more expressive deep learning models. Comparisons with deep multi-view learning methods would help establish its effectiveness on complex real-world tasks.

  4. The theoretical analysis is insufficient, particularly regarding convergence and generalization. Although the paper establishes linear function-class preservation and discusses computational complexity, it provides limited theoretical analysis of the overall learning framework. In particular, convergence guarantees for the joint non-convex optimization and generalization bounds are missing. Further theoretical investigation would strengthen the reliability and understanding of the proposed method.

Questions for authors
  1. The learned simplex gates exhibit substantially different allocation patterns across datasets. For example, on SCENE, nearly 97% of the structural weight is assigned to a single view, while the remaining views receive almost zero weight. Could the authors further explain the implications of such highly concentrated gate allocations? Could this behavior potentially prevent useful shared information from other views from being fully utilized?

  2. The current experiments mainly consider datasets with complete multi-view features and label information. In real-world applications, some views may be missing, noisy, or of inconsistent quality. How robust is SG-TPR under such conditions? In particular, can the adaptive gating mechanism automatically reduce its reliance on a view when the quality of its information significantly deteriorates?

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

This paper studies multi-view multi-label feature selection and proposes SG-TPR, which organizes each view's linear mapping into a residual branch and a low-rank structural branch with a shared label factor. The same label factor participates in the residual transformation, structural prediction, signed label-correlation calibration, and feature-row sparsity regularization. A simplex gate adjusts the strength of structural regularization across views. The authors show that this parameterization preserves the full class of linear mappings, and therefore attribute its gains to changes in regularization and optimization geometry rather than increased expressivity. Experiments on seven datasets use common feature budgets and a downstream classifier, with structural ablations, an independently retuned uniform-gate control, and statistical analysis.

Critical strengths
  1. The same label factor connects structural prediction, residual transformation, and regularization, making the design coherent. Eq. (5) and Appendix B show that the full linear mapping class is preserved, while the gates are interpreted as adaptive regularization. This gives the contribution a clear scope.

  2. The seven datasets use common splits, feature budgets, and a classifier, with preprocessing and tuning restricted to training data. The method achieves the best average rank on all four metrics and retains a mean AP advantage over an independently retuned uniform-gate control. The authors also state the limitations of the statistical tests.

Critical weaknesses
  1. Eq. (4) is a bilinear identity, so the main methodological addition lies in the coupled regularization through the label factor. However, the key Additive-LR control inherits Full's hyperparameters (Figure 3 and Table 4), unlike the independently retuned Uniform-Gate control. SG-TPR's advantage over a properly tuned standard additive low-rank model therefore remains uncertain.

  2. NR-CILL, MSFS, and other baselines in Tables 6–7 use reimplementations with approximations or fixed implementation choices, whose fidelity needs clarification. Given the small AP gaps on some datasets, implementation differences cannot yet be ruled out as a factor in the rankings.

Questions for authors
  1. Are independently tuned Additive-LR results available? If not, please clarify the scope of the advantage supported by the current comparison.

  2. Under what data or label structures should residual coupling through the shared label factor be more effective? Can the existing analysis and dataset-level results help explain these conditions?

  3. What key approximations were used in the NR-CILL and MSFS reimplementations, and what checks support consistency with the original optimization procedures and feature rankings?

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

Related papers

Discussion (0)

Sign in to comment.