acceptodds
Under review as a conference paper at ICLR 2027

HyperANFIS: Enhancing Rule Representation and Interpretability in Adaptive Neuro-Fuzzy Systems via Hyperbolic Geometry

Abstract

The adaptive neuro-fuzzy inference system (ANFIS) is an interpretable reasoning framework capable of generating explicit IF-THEN fuzzy rules, making it suitable for tasks requiring transparent reasoning. However, existing ANFIS models generally construct rule antecedents and perform inference in Euclidean space, limiting their representational capacity and predictive performance. To address this issue, we propose Hyperbolic ANFIS (HyperANFIS), a hyperbolic extension of ANFIS. HyperANFIS preserves the fuzzy semantics and core architecture of conventional ANFIS while performing rule-prototype learning, rule activation, and consequent aggregation in hyperbolic space. It also retains the ability to generate interpretable IF-THEN rules. By exploiting the representational properties of hyperbolic geometry, HyperANFIS strengthens the fuzzy inference process, thereby improving predictive accuracy, inter-rule collaboration, and the credibility of its interpretable rules. Experimental results show that HyperANFIS consistently outperforms the standard ANFIS baseline and various ANFIS variants across all datasets, while also generating higher-quality fuzzy rules.

Primary area:
interpretability and explainable AI
Keywords:
Hyperbolic Geometry, ANFIS, IF-THEN Fuzzy Rules, Interpretability

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 4 reviews, above 27% of the others.
Reviewer_JMEXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes HyperANFIS, a hyperbolic neuro-fuzzy model combining geodesic prototype-based antecedents, local consequent mappings, and Fréchet aggregation. It motivates negative curvature as a way to organize rule representations, alleviate rule crowding, and improve predictive performance and rule quality. Experiments on five classification datasets compare the model with ANFIS variants, while analyses of rule activation, class coverage, and prototype organization support the proposed interpretation of more cooperative and trustworthy inference.

Critical strengths
  1. A coherent geometric formulation of neuro-fuzzy inference. The antecedent memberships, local consequents, and aggregation are integrated into a consistent manifold-based architecture. This provides a concrete algorithmic contribution and a useful framework for investigating how geometry affects rule-based prediction.
  2. Encouraging predictive results with explicit analysis of rule behavior. The experiments across five datasets report promising predictive performance, and the paper examines how rules are used in addition to reporting aggregate accuracy. These results make the method potentially useful and provide empirical motivation for studying its proposed geometric mechanism.
Critical weaknesses
  1. The central geometric contribution is not sufficiently established. The feature-combination tree in Figure 1 describes an enumeration whose grouping depends on feature order; it does not, by itself, establish a task-relevant semantic hierarchy. Likewise, the growth in the number of rules, R = product_j M_j, arises from enumerating feature-wise combinations. A fixed set of global prototypes can avoid this growth in Euclidean space as well.
  2. HyperANFIS changes the antecedent construction from coordinate-wise membership products to geodesic prototype memberships. Consequently, comparisons with conventional ANFIS do not isolate the contribution of negative curvature, even when rule counts are matched.
  3. Appendix B.1, Eq. (39), already defines a Euclidean counterpart retaining the rule count, consequent parameterization, and normalized activation, but Table 1 does not report its results. This omission matters because the paper attributes its benefits to the geometry of rule organization.
  4. The comparative rule-quality claim also remains insufficiently established: Section 4.3 evaluates rule quality only against traditional ANFIS, although Table 1 includes FSRE-AdaTSK, which directly addresses feature selection and rule extraction. The current evidence therefore leaves both the source of the geometric benefit and the added value of the resulting rule representations insufficiently resolved. These gaps concern the paper's central contribution and materially limit my assessment.
  5. The rule-quality analysis does not establish more trustworthy reasoning. Figure 3 reports increases in activation similarity from 0.006 to 0.432 and in the effective number of active rules from 1.14 to 6.67. These quantities measure activation overlap, but do not establish that the additional active rules provide useful, complementary information. If all rule consequents for a given input coincide at q, their Fréchet aggregate remains q for any normalized activation weights. The effective rule count can then increase from nearly 1 to R without changing the prediction. This counterexample limits the interpretation of the statistics; it does not assert that the trained model actually contains duplicate rules.
  6. Furthermore, Table 2 and Figure 3(b) assign benign/malignant labels to rule prototypes and identify incorrectly learned prototypes without a clear operational criterion. Antecedent centers, input-dependent consequents, and output class prototypes are different mathematical objects. Labeling a rule by its isolated output, the complete model's prediction at its center, or the labels of the samples it covers can lead to different diagnoses. Thus, neither the overlap statistics nor the prototype-level analysis currently establishes the claimed improvement in meaningful rule cooperation and class representation. Together with the unresolved geometric contribution, this leaves the central conclusions insufficiently supported and places my recommendation slightly below the acceptance threshold.
Questions for authors
  1. Evidence for the geometric and comparative contribution. Can you provide the direct comparison with the Euclidean counterpart in Appendix B.1, Eq. (39), under matched rule counts, consequent order, data splits, and reasonably comparable training and tuning budgets? Can the claimed rule-quality advantage also be assessed against a method that directly improves rule structure, such as FSRE-AdaTSK? Such evidence could change my assessment of the geometric contribution and its added value.
  2. Clarification of the rule-quality evidence. How are fixed benign/malignant rule labels and incorrectly learned prototypes defined and computed? What evidence shows that secondary active rules contribute complementary information that matters to the final decision? Clear definitions and evidence linking rule cooperation to the prediction could resolve the second concern.
  3. Clarification of metric definitions and the baseline evaluation pipeline. Table 1 reports Accuracy = 0.9035 and Recall = 0.9712 for FCM-ANFIS on WDBC. The supplied HyperANFIS/evaluate.py computes macro-averaged recall, but main.py runs only HyperANFIS, and the supplementary material does not establish that the FCM-ANFIS result used the same evaluator. For binary classification, metrics computed from the same predictions satisfy A >= 2 * R_macro - 1; the bound also holds for identically weighted averages over the same runs. If the reported Recall is macro-averaged, it would require A >= 0.9424, inconsistent with 0.9035. If it is positive-class recall, the reported metrics can be mathematically feasible. Please specify the definition used for each method and provide the actual FCM-ANFIS evaluation pipeline and per-seed confusion matrices or predictions. This is a conditional consistency concern: a consistent definition supported by the corresponding results would resolve it.
Reviewer_HmDDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes HyperANFIS, which extends traditional ANFIS by introducing hyperbolic geometry for rule representation, activation, and aggregation. The method aims to improve prediction accuracy while maintaining interpretable IF–THEN rules. Experiments on five classification datasets report better performance than conventional ANFIS and several variants. The authors also analyze learned rules on the WDBC dataset to support their interpretability claims.

Critical strengths
  1. The paper provides a coherent extension of ANFIS to hyperbolic space, covering rule representation, activation, and consequent aggregation within a unified framework.

  2. The evaluation covers five classification datasets, and the WDBC case study provides concrete examples of learned fuzzy rules. These analyses are useful for understanding the model's behavior, although the interpretability claims need stronger evidence.

Critical weaknesses
  1. The contribution of hyperbolic geometry is not established. Appendix B.1 defines a geometry-matched Euclidean variant, but its results are not reported. Since HyperANFIS also changes the rule representation and inference architecture, comparisons with conventional ANFIS cannot isolate the benefit of hyperbolic geometry. This weakens the paper's main novelty claim.

  2. Concerns about experimental reliability and interpretability. Table 1 reports 0.9035 accuracy and 0.9712 recall for FCM-ANFIS on WDBC, which are inconsistent if recall is macro-averaged. The supplementary code also lacks the baseline implementations and multi-seed experiments needed to reproduce the main results. Furthermore, the interpretability analysis mainly examines rule activation and coverage, without showing whether the extracted IF–THEN rules faithfully explain the predictions.

Questions for authors
  1. Could the authors report results for the geometry-matched Euclidean variant defined in Appendix B.1? This comparison is essential to determine whether the improvements come from hyperbolic geometry or the redesigned rule architecture.

  2. Could the authors clarify the recall definition in Table 1, particularly the WDBC results for FCM-ANFIS? Could they also provide per-seed results and the scripts used to generate the comparison table?

  3. How faithfully do the simplified IF–THEN rules in Table 2 reflect the actual model predictions? Could the authors quantify this rather than relying mainly on rule activation and coverage statistics?

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

The paper introduces HyperANFIS, a neuro-fuzzy model that represents rules with hyperbolic prototypes, computes rule activations from geodesic distances, and aggregates rule consequents on a hyperbolic manifold. It reports improved classification performance over several ANFIS variants on five datasets and uses a WDBC case study to argue that the learned rules are more interpretable and better organized.

Critical strengths
  1. The geometric formulation is coherent. Hyperbolic geometry is incorporated throughout rule matching, local consequent construction, and aggregation, giving the model a consistent inference pipeline. This is an interesting way to extend the representation of fuzzy rules beyond conventional feature-wise antecedents.
  2. The reported results show improvements over several neuro-fuzzy baselines across all five datasets, including substantial gains on Zoo and WDBC. The WDBC case study also gives concrete examples of the learned rules and connects their representative features to recognizable diagnostic patterns, making the proposed model easier to inspect.
Critical weaknesses
  1. The role of negative curvature is difficult to separate from the other design changes. HyperANFIS replaces feature-wise antecedents with global prototypes and changes the consequent aggregation and prediction mechanism. The performance gains support the complete architecture, while the paper offers limited insight into how much of the benefit comes from geometry and which task structures favor it. This limits the general lessons that can be drawn from the empirical results.
  2. The interpretability argument is less developed than the predictive evaluation. The WDBC analysis presents plausible feature patterns, broader rule participation, and more balanced class representation. Following a prediction also involves understanding global geodesic antecedents and the aggregation of several rule consequents. These computations make it important to explain how the displayed IF-THEN summaries help a reader trace a model decision. That connection receives relatively little attention, leaving the practical interpretability advantage over classical ANFIS unclear.
  3. Sections 2.1 and 2.2 are near duplicates. I believe this is a minor writing issue and can be easily fixed.
Questions for authors
  1. How do the authors view the relative contributions of the geometric representation and the other architectural changes? What characteristics of a task make this geometric formulation particularly useful?
  2. Could the authors use an example to explain how the displayed IF-THEN rules connect to an actual prediction and what interpretive advantage they offer over classical ANFIS?
Reviewer_21WFOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

HyperANFIS maps inputs and rule prototypes into hyperbolic space, derives rule activation from geodesic distance, constructs intrinsic consequents, and aggregates them through an approximate weighted Frechet mean. It reports classification results on five tabular datasets and a qualitative rule analysis on the WDBC breast-cancer dataset.

Critical strengths

The paper specifies a complete geometric pipeline rather than adding a single embedding layer.

Critical weaknesses

The central comparison confounds geometry and architecture. HyperANFIS uses global radial rules, while conventional ANFIS uses coordinate-wise memberships; the paper states that its zero-curvature counterpart is not conventional ANFIS (Appendix D.2). An architecture-matched Euclidean baseline is essential. The claimed hierarchical advantage is not established for most selected datasets. No analysis shows that their structure favors negative curvature. There are no decisive ablations for curvature, radial versus coordinate-wise rules, Frechet versus Euclidean aggregation, or the auxiliary training losses. The table reports means without uncertainty, which is important for small datasets. The interpretability claim rests on selected WDBC rules and activation statistics from a fixed seed. More active or correlated rules do not by themselves imply greater fidelity, stability, or human understanding. Three Karcher iterations are also not shown to approximate the mean accurately.

Questions for authors

How does HyperANFIS compare with its exact Euclidean radial-rule counterpart under matched parameters, rules, and optimization?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.