acceptodds
Under review as a conference paper at ICLR 2027

SCAMFS: State-level Causal Asymmetry Modeling for Multi-label Feature Selection

Abstract

Multi-label feature selection aims to identify compact and informative feature subsets from data with multiple correlated labels. Existing methods typically evaluate feature relevance at the label-variable level, treating each binary label as an indivisible unit. This may obscure state-specific patterns, especially when rare positive states are overwhelmed by dominant negative states, resulting in state-level dilution and label blocking. To address this issue, we propose SCAMFS, State-level Causal Asymmetry Modeling for Multi-label Feature Selection, as a causally motivated state-level feature selection framework. SCAMFS introduces a state-level decomposition strategy that decomposes each label into state-specific events, and further defines State-Specific Mutual Information (SSMI) and Differential State Mutual Information (DSMI) to model feature–state relevance and cross-label state dependencies. It also employs an explanatory competition mechanism to recover informative features masked by correlated label states. Experiments on seven real-world datasets show that SCAMFS outperforms representative baselines on most metrics. The anonymous code repository is available at: https://anonymous.4open.science/r/SCAMFS-code-0C33.

Primary area:
causal reasoning
Keywords:
Feature selection, multi-label, causal asymmetry, state-level decomposition, explanatory competition.

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_bfkyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes SCAMFS, a state-level causal asymmetry modeling framework for multi-label feature selection. The method decomposes binary labels into state-specific events (presence/absence) and introduces SSMI and DSMI metrics within a four-stage pipeline covering cross-label dependency modeling, local state structure discovery, feature recovery, and reassessment. The approach is evaluated on seven standard multi-label datasets against several baselines.

Critical strengths

1.The conceptual shift from variable-level to state-level decomposition, treating label presence and absence as distinct events, offers a theoretically interesting perspective on multi-label feature selection. 2.The evaluation covers seven standard multi-label datasets and compares against eight baseline methods, providing a reasonably broad empirical testing ground despite the mixed results.

Critical weaknesses

1.The manuscript suffers from chaotic organization that hinders comprehension, making the method difficult to follow. 2.The Related Work section fails to capture the state-of-the-art, with only one 2025 citation and heavy reliance on literature before 2024. It ignores significant recent advances in causal multi-label feature selection and deep learning-based approaches. The discussion of prior work is merely a list of methods without a critical synthesis of their shortcomings, failing to establish a compelling gap that SCAMFS fills. 3.The ablation study (Figure 2) shows that Phases 1 and 2 contribute little to the final performance on most datasets, yet the paper provides only a brief justification for retaining them without deeper analysis. Furthermore, there is no independent ablation of SSMI versus DSMI, nor a direct comparison between state-level decomposition and traditional variable-level modeling to validate the core hypothesis.

Questions for authors

1.State-level decomposition costs more computation but only gives marginal accuracy gains. How do you justify this trade-off? 2.Definitions 2 and 3 use a conditioning set ZZ, but it's not clear what variables it can contain or how you actually pick it. Can you clarify? 3.SCAMFS has six parameters. The sensitivity analysis only covers four of them, what about ss and hh?

Reviewer_weudOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses the challenges posed by cross-label dependencies and imbalanced label-state distributions in multi-label feature selection and proposes a state-level feature selection method, SCAMFS. By separately analyzing label absence and presence, the method uses SSMI and DSMI to characterize feature–label and cross-label dependencies, respectively. It further introduces an explanatory competition mechanism to recover features whose contributions are masked by correlated labels, followed by redundancy elimination and validity checking.

Critical strengths
  1. The motivation is clear, and the problem has practical relevance. Considering label correlations alone in multi-label settings may overlook information associated with rare label states. The paper provides an in-depth investigation of this issue.
  2. By separately analyzing label presence and absence, the method captures distributional differences across states and provides a finer-grained characterization of asymmetric cross-label associations. This perspective is insightful.
  3. The paper explicitly revisits features whose contributions are masked by label correlations, helping to alleviate label blocking.
  4. The experimental evaluation is relatively comprehensive, including comparisons on seven datasets from different domains, parameter sensitivity analysis, phase-wise ablation studies, and computational cost analysis. These results support the effectiveness of the method.
Critical weaknesses
  1. SSMI is formulated as the KL divergence between a conditional distribution and the marginal distribution. The authors are encouraged to clarify how it differs from existing related information measures.
  2. Cross-state filtering uses a multiplicative threshold of 1.2, and the recovery procedure permits only one successful recovery per target state. Further justification for these choices and an analysis of their effects on the results would be helpful.
  3. A feature being more strongly associated with another label state does not necessarily imply that it lacks value for the current state. This point warrants further analysis and clarification.
  4. Rare states typically contain few samples. Under such conditions, how can useful state-specific information be reliably distinguished from small-sample noise?
Questions for authors
  1. Is the empirical effectiveness of the method primarily attributable to state-level evaluation or to the explanatory competition mechanism?
  2. After merging the feature subsets selected for individual label states, does the resulting feature set still contain a degree of redundancy?
  3. If multiple features were allowed to be recovered for each target state, how would predictive performance and computational cost change?
Reviewer_t9iBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes SCAMFS, a state-level causal asymmetry modeling framework for multi-label feature selection. The method decomposes binary labels into presence and absence states and introduces State-Specific Mutual Information (SSMI) and Differential State Mutual Information (DSMI) to capture feature–state relevance and asymmetric cross-label dependencies. It further incorporates explanatory competition to recover informative features masked by correlated label states, followed by redundancy elimination to refine the selected feature subset. The authors evaluate SCAMFS on seven real-world multi-label datasets against seven representative baselines using multiple classification metrics. The reported results demonstrate competitive or superior performance across most datasets and metrics, supported by ablation studies, parameter sensitivity analysis, runtime comparisons, and statistical significance tests.

Critical strengths

The paper addresses a meaningful limitation of existing multi-label feature selection methods by explicitly modeling state-specific feature relevance and asymmetric label dependencies. The proposed framework is well-structured, and the experimental results across seven datasets demonstrate consistent improvements over representative baselines, supporting the practical value of the approach.

Critical weaknesses

The theoretical justification for the causal interpretation remains insufficient, as the proposed information measures primarily capture statistical dependencies rather than causal relationships. More importantly, inconsistencies between the reported Coverage results and its formal definition raise concerns about experimental reliability. The lack of controlled comparisons under matched feature-selection budgets and unclear details in the phase-wise ablation study further weaken the empirical conclusions.

Questions for authors
  1. Could the authors clarify the theoretical assumptions supporting the causal interpretation of SSMI, DSMI, and the state-specific Markov blankets? In particular, how does the proposed framework distinguish causal relationships from statistical dependencies?
  2. The Coverage values in Table 6 appear inconsistent with the unnormalized definition in Eq. (24). Could the authors clarify whether normalization was applied and confirm that all methods were evaluated consistently?
  3. How is the Phase 1 performance in Fig. 2 evaluated, given that this phase does not directly produce a selected feature subset? Additionally, could the authors clarify whether the compared methods select similar numbers of features, or provide results under matched feature-selection budgets? I am open to reconsidering my overall rating upward if the authors can address these concerns during the rebuttal.
1 comment

Dear Reviewer,

Thanks for your time and effort in reviewing this paper. Your comments are valuable to improve the quality of this paper. As the author–reviewer discussion period will end in about 1 day, we would be grateful if you could take another look at our rebuttal when convenient and let us know whether it has adequately addressed your concerns.

Thanks again for your time and help in improving the quality of this paper.

Regards from the authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.