Explainability Constraints in X-ANFIS

• The paper introduces a bi-objective optimization framework that alternates gradient steps on prediction and explainability losses to recover non-convex Pareto optimal solutions.
• It employs structural constraints such as Cauchy membership functions and sparsity regularization to ensure ordered, semantically clear, and scalable fuzzy rule sets.
• By integrating explicit uncertainty quantification and domain-specific physics-informed vetoes, X-ANFIS enhances transparency and auditability in risk-sensitive applications.

Adaptive Neuro-Fuzzy Inference Systems (ANFIS) have evolved into powerful explainable models through the integration of explicability constraints at multiple levels of architecture and training. Explainability-Optimized ANFIS (X-ANFIS) systems formalize and enforce constraints to ensure transparent, auditable decision processes while controlling the trade-off between predictive accuracy and semantic interpretability. The main approaches to explainability constraints in X-ANFIS include bi-objective optimization, antecedent space sculpting, dual sparsity regularization, uncertainty quantification, physics-informed hard rules, and additive neuro-symbolic architectures.

1. Bi-Objective Explainability Optimization in X-ANFIS

The central paradigm in modern X-ANFIS is the explicit bi-objective minimization of predictive loss and explainability loss, typically formalized as: $\min_\theta \quad \{ L_P(\theta), L_E(\theta) \}$ with $L_P(\theta)$ denoting mean-squared prediction error and $L_E(\theta)$ quantifying deviation from a fixed semantic distinguishability target $D_\text{target}$. In "Alternating Bi-Objective Optimization for Explainable Neuro-Fuzzy Systems" (Khaled et al., 22 Feb 2026), explainability is enforced not via convex scalarization, but by alternating explicit gradient steps on $L_P$ and $L_E$.

The explainability loss is precisely: $$L_E(\theta) = \sum_{f=1}^m \sum_{(i,j)\in J_f} \tfrac{1}{2}\left(D_{ij}^{(f)}(\theta) - D_\text{target}\right)^2,$$ with $D_{ij}^{(f)}(\theta)$ the distinguishability between adjacent fuzzy sets parameterized by centers $c$ and scales $\gamma$: $L_P(\theta)$0 Crucially, only centers $L_P(\theta)$1 are updated during the explainability (X-)pass to avoid scale degeneracies. The decoupling of loss updates allows X-ANFIS to recover non-convex regions of the accuracy–explainability Pareto front, which are inaccessible to classic weighted-sum methods.

2. Structural and Gradient Constraints for Semantic Clarity

Training stability and semantic clarity are enforced via both membership function design and initialization constraints. X-ANFIS employs Cauchy membership functions for their favorable gradient properties: $L_P(\theta)$2 FCM-based semantically controlled initializations are utilized: features are clustered, and scales $L_P(\theta)$3 are set so that adjacent clusters conform to the target distinguishability $L_P(\theta)$4. During optimization, $L_P(\theta)$5 is frozen on the X-pass; $L_P(\theta)$6 is shifted to minimize distinguishability error, and gradient clipping is applied to constrain update magnitudes.

3. Sparse Rule Sets, Additive Indexing, and Linear Scalability

Classical ANFIS scales poorly: $L_P(\theta)$7 rules for $L_P(\theta)$8 features and $L_P(\theta)$9 fuzzy sets per input, impeding both interpretability and manageability. Additive index models and feature projection architectures, as in FAME ("Fuzzy Additive Models for Explainable AI") (Gokmen et al., 9 Apr 2025) and KANFIS (Yong et al., 3 Feb 2026), recast ANFIS in terms of one-dimensional SISO subnetworks or additive aggregation:

• FAME enforces hard constraints (coupled centers and shared boundary widths) on two-sided Gaussian MFs, ensuring per-feature fuzzy labels are ordered, have minimal overlap, and never fire more than two rules per axis.
• KANFIS replaces product-based rule firing with additive aggregation and entropy-based sparsity regularization. A soft-mask matrix $L_E(\theta)$0 enforces that each rule uses at most a few inputs, with an additional distinctiveness penalty to maintain unique rule semantics:

$L_E(\theta)$1

These strategies yield a linear scaling with feature count, maintain interpretability, and empirically reduce the number of active rules without substantial loss in accuracy.

4. Attribute and Rule-Level Sparsity Penalties

Explainability is further supported by explicit regularization of both attribute and rule weights: $L_E(\theta)$2 where $L_E(\theta)$3, $L_E(\theta)$4, and $L_E(\theta)$5 is the sigmoid. The ADAR framework (Liu et al., 27 Apr 2025) applies these penalties along with automated growth and pruning. Attribute- and rule-level weights are learned end-to-end, dynamically adapt importance, and prune irrelevant features/rules when their post-sigmoid scores drop below thresholds. Practical implications include:

• Drastic reduction in model overlap (Overlap Index) and improved separation of fuzzy sets (Set-Position Index).
• Compact final rule bases (e.g., PM2.5 prediction: from 12 to 9 rules, pruning 40% of feature-rule pairs) with only minor accuracy loss.

5. Explicit Uncertainty Quantification and FOU Constraints

Interval Type-2 X-ANFIS (Khaled et al., 26 Jan 2026, Yong et al., 3 Feb 2026) extends explainability to uncertainty by associating each fuzzy set with a Footprint of Uncertainty (FOU): $L_E(\theta)$6 Upper and lower Gaussian MFs define the FOU, and minimum width and ordering constraints ensure that fuzzy sets remain semantically meaningful. Feature-level $L_E(\theta)$7, rule-level $L_E(\theta)$8, and instance-level $L_E(\theta)$9 interval widths decompose prediction uncertainty, tracing ambiguity back to its structural source. Constraints such as fixed $D_\text{target}$0, clipped gradients, and bounded type-reduction factors maintain comprehensible, stable uncertainty semantics, supporting decision auditability in risk-sensitive domains.

6. Domain-Specific and Physics-Informed Constraints

X-ANFIS can integrate domain-specific, physically meaningful hard constraints to enforce explainability. In quantum error attribution (Hassan et al., 22 Feb 2026), the Bhattacharyya veto is employed: $D_\text{target}$1 with $D_\text{target}$2 the Bhattacharyya distance between measured and ideal output distributions. This constraint, grounded in the Data Processing Inequality for CPTP channels, forbids the attribution of errors to hardware noise when the divergence exceeds physically allowable thresholds. The veto is neither a regularizer nor learnable: it is a hard override—improving transparency and providing an auditable logical basis for every classification and failure mode.

A "safe-failure" mode is added, with the model flagging uncertain cases for manual review where the predicted probability lies within a margin (e.g., $D_\text{target}$3), further supporting reliable and explainable deployment.

7. Empirical Performance and Trade-off Analysis

Empirical studies on standard regression and diagnostic problems (Khaled et al., 22 Feb 2026, Gokmen et al., 9 Apr 2025, Liu et al., 27 Apr 2025, Khaled et al., 26 Jan 2026, Hassan et al., 22 Feb 2026, Yong et al., 3 Feb 2026) consistently show that:

• X-ANFIS architectures recover high-distinguishability solutions ($D_\text{target}$4) at negligible predictive performance loss (1–3% in $D_\text{target}$5).
• Traditional multi-objective (MO-ANFIS) scalarization is limited to the convex hull of the Pareto front and must sacrifice 10–20% in $D_\text{target}$6 to reach similar distinguishability.
• ADAR and FAME architectures achieve efficient, interpretable performance on high-dimensional data, mitigating exponential rule growth.
• Interval Type-2 systems provide robust, interpretable uncertainty intervals with low model variance.
• Physics-vetoes prevent unexplained or logically impossible attributions, raising user trust in critical settings.

The following table synthesizes the principal mechanisms and their explainability roles:

Mechanism	Constraint Type	Explainability Effect
Bi-objective alternation	Optimization	Non-convex Pareto points; explicit targets
Sculpting (FAME, hard anchor)	Structural/Hard	Ordered, distinct fuzzy labels; few firings
Attribute/rule sparsity (ADAR)	Regularization/Pruning	Small, active rule base; irrelevance removal
Interval type-2 MFs (IT2-ANFIS)	Bounded, interval sets	Traceable multi-level uncertainty
Additive aggregation (KANFIS)	Structural/Soft-masked	Linear scaling, few features per rule
Physics-informed veto	Hard logical rule	Auditable, domain-valid inference

8. Limitations and Best Practices

Persistent limitations include:

• Gaussian MFs may be restrictive under non-Gaussian data distributions (Khaled et al., 26 Jan 2026).
• Epistemic but not aleatoric uncertainty is captured in current interval type-2 architectures.
• Overly aggressive pruning may underfit, while lax regularization can reintroduce ambiguity and rule overlap.
• Domain-specific vetoes require prior domain knowledge and cannot be universally automated.

Best practices for maintaining X-ANFIS interpretability include stratified initialization, rule-count capping, gradient clipping, fixed or bounded type-reduction factors, and explicit monitoring of overlap and position indices for fuzzy sets (Khaled et al., 22 Feb 2026, Liu et al., 27 Apr 2025, Khaled et al., 26 Jan 2026).

• X-ANFIS bi-objective paradigm and alternation: (Khaled et al., 22 Feb 2026)
• FAME/FAM antecedent space sculpting: (Gokmen et al., 9 Apr 2025)
• Physics-informed veto in quantum diagnostics: (Hassan et al., 22 Feb 2026)
• Interval Type-2 uncertainty and structural constraints: (Khaled et al., 26 Jan 2026)
• Dual sparsity and pruning (ADAR): (Liu et al., 27 Apr 2025)
• Additive neuro-symbolic masking and scalability (KANFIS): (Yong et al., 3 Feb 2026)