acceptodds
Under review as a conference paper at ICLR 2027

Antecedent-Aligned Rule Transfer with Partial-Overlap Consequent Fusion for Multi-Center TSK Fuzzy Classification

Abstract

Multi-center Takagi–Sugeno–Kang (TSK) fuzzy classification faces challenges arising from asymmetric sample sizes and partially overlapping local rule bases, as independently learned antecedents may not denote comparable fuzzy regions across centers. We propose MCRT-TSK, which establishes shared antecedent identities for semantic rule matching while allowing each center to retain only locally supported rules. At the target center, a sample-count-adaptive criterion combines source-side support with target activation evidence, and consequent blocks are fused only from sources containing the matched antecedent. A rule-level prior and a response-level anchor regularize a convex target consequent-estimation problem with a closed-form solution. On 18 UCI datasets, MCRT-TSK achieves the highest or tied-highest target-test ACC among methods with valid outputs on 16 of the 18 datasets. Across three stress-test families, shared antecedent transfer yields positive mean ACC, Macro-F1, and G-mean gains over local-target training; identity-preserving consequent transfer also yields positive mean margins over a within-order shuffled-prior control in all nine task-family–metric summaries. Statistical, ablation, efficiency, and rule-level evidence supports the full MCRT-TSK under the simulated-center protocol and aligned rule transfer in the stress tests.

Primary area:
interpretability and explainable AI
Keywords:
Takagi–Sugeno–Kang (TSK) fuzzy systems, Multi-center learning, Transfer learning, Rule alignment, Knowledge fusion

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

This paper introduces MCRT-TSK, a multi-center TSK fuzzy classification framework that transfers knowledge by explicitly aligning rule antecedent identities across centers, while allowing each center to retain only locally supported rules. It proposes (i) a shared antecedent identity space defined by the main center, (ii) a support-aware, sample-count-adaptive rule selection at the target that mixes source support with target activation, and (iii) a partial-overlap consequent fusion that aggregates only from sources retaining the matched antecedent, followed by a closed-form convex target estimation. Experiments on 18 UCI datasets verify its effects.

Critical strengths
  1. Reasonable method design: Explicit antecedent identity keys (feature subset plus fuzzy-set indices) provide a principled transfer interface for TSK rules under partial rule overlap, which is not enforced in prior multi-center fuzzy/TSK approaches.

  2. Comprehensive evaluation: Comprehensive evaluation across 18 UCI datasets with 10 runs, plus controlled heterogeneity stress tests and natural multi-center Heart cohorts; multiple metrics (ACC, Macro-F1, G-mean) are reported.

Critical weaknesses
  1. Antecedent identity does not establish consequent comparability. Although matched rules share the same unnormalized membership function, their normalized firing strengths depend on each center’s retained rule set. Consequently, the associated coefficients are estimated under different design matrices and potentially different feature correlations. Identical antecedent keys therefore do not, by themselves, justify directly averaging or regularizing consequent blocks. This paper does not establish when such parameter transfer is valid rather than merely enforcing syntactic correspondence.

  2. Support compatibility is not predictive compatibility. The fusion weights depend on unlabeled activation profiles, rather than agreement between source and target class responses within matched regions. Sources with similar activation statistics but conflicting conditional label distributions can therefore receive similar weights. The response-level anchor may additionally impose inappropriate source class tendencies under label-prior shift. The heuristic weighting is acknowledged, but its ability to identify harmful transfer is insufficiently evaluated.

  3. The stress tests do not clearly show that consequent transfer improves performance. Although the full method outperforms the shuffled-prior control on average, none of these differences is statistically significant in Table S15, and gains occur in only two of four Heart sites. The simpler Shared-Antecedent model, which does not transfer consequents, also ranks first on most Heart sites. Thus, the results show that correctly matched priors are preferable to shuffled ones, but do not clearly establish that transferring consequents is better than learning them from target data alone.

Questions for authors
  1. Under what conditions are consequent coefficients transferable when the same antecedent has different normalized firing strengths across centers? Can you compare coefficient fusion against response-based transfer evaluated on a common target design matrix?

  2. How does the weighting distinguish sources with similar support profiles but contradictory class responses? Can you evaluate controlled conditional-label mismatch and compare against target-supervision-based compatibility weighting?

  3. Can you report paired Full-versus-Shared-Antecedent results across all stress settings, with uncertainty and significance tests? For the Heart comparison, can preprocessing be matched across configurations to separate antecedent-transfer gains from the benefit of pooled training statistics?

Reviewer_y4FCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MCRT-TSK, a rule-transfer method for multi-center TSK fuzzy classification. It addresses the problem that antecedent semantics are not comparable across centers when sample sizes are asymmetric and local rule bases only partially overlap. The method consists of three parts: (1) the main center defines a shared antecedent-identity space, which serves as the key for cross-center rule matching. (2) auxiliary centers receive main-center consequent priors only for identities matched with the main center. (3) the target center selects rules by adaptively mixing source-side and local evidence according to its sample size, fuses consequents only from sources that retain the same identity, and solves the resulting problem in closed form. Experiments show that the method achieves the best or tied-best accuracy on 16 of 18 UCI datasets, and yields consistent positive gains over local training and over a shuffled-prior control in the stress tests.

Critical strengths
  1. The shuffled-rule-prior control is well designed, it shows that the gains come from identity-aligned consequent reuse, rather than from merely introducing source data.
  2. The whole pipeline admits closed-form solutions and is orders of magnitude faster than the transfer baselines, while still preserving rule-level interpretability.
Critical weaknesses
  1. The multi-center setting of the main benchmark is simulated by stratified random splitting, so the centers are approximately IID. A negative gain appears only in the single natural multi-site scenario (Heart), and the authors do not analyze the reason.
  2. Key design choices, such as the fusion weights and the temperature, are heuristic and insufficiently justified. In addition, the strongest baseline is excluded on multiple datasets due to timeouts, which weakens the main claim.
Questions for authors
  1. The antecedent-identity space grows combinatorially with the feature dimensionality. Is the method still feasible on high-dimensional data?
  2. When the overlap among source rules is very low, does the method degrade to the point of being worse than training locally on the target data?
Reviewer_UQRgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a multi-center TSK fuzzy classification framework designed to handle partially overlapping local rule bases and asymmetric sample allocation. The method designs shared antecedent identities, performs support-aware rule selection, and fuses matched consequent parameters for target-center learning. Experiments demonstrate competitive performance against existing approaches.

Critical strengths

1.The paper addresses a clearly defined problem, which is the difficulty of transferring knowledge across partially overlapping rule bases in multi-center TSK fuzzy classification. 2.The proposed framework is technically coherent. 3.The paper is generally well organized, and the mathematical formulations are clearly presented.

Critical weaknesses
  1. The main experiments do not fully test cross-center transfer. Most centers are created by randomly splitting the same dataset, so the improvements may largely come from using more training data rather than the proposed rule-alignment mechanism. Although additional heterogeneity tests are included, stronger comparisons under genuinely different center distributions would be helpful.
  2. There is no pooled-data TSK, source-only model, independent-partition + consequent-averaging baseline, non-fuzzy baseline (such as XGBoost, tuned MLP), or plain biased regularization toward a source prior.
  3. Some baseline results and evaluation details need clarification. For example, 1-TSK performs poorly on certain datasets, and some target test sets are very small, making accuracy differences difficult to interpret. Could the authors explain these results and clarify how hyperparameters were selected without using target test data?
  4. The mechanism evidence is thin and over-claimed. Shared-Antecedent excludes source consequents, and the full vs. Shuffled-Rule-Prior result is reported only as means over 4–6 blocks without p values or CIs, yet is called consistent evidence.
Questions for authors
  1. Can you add pooled-data, source-only, independent-partition + averaging, and non-fuzzy baselines (such as XGBoost, tuned MLP)?
  2. Were hyperparameters and ablation variants tuned without using target test data?
  3. Can you discuss p-values or CIs for full vs. Shuffled-Rule-Prior and Full vs. Local-Target in the main text?
  4. Does the method still help on genuinely heterogeneous multi-site data, including conflicting sources?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.