Fuzzy Rule-Based Model Explanations

• Fuzzy rule-based model explanations are a framework that translates fuzzy inference systems into interpretable linguistic constructs using fuzzy sets, membership functions, and IF-THEN rules.
• Algorithmic approaches such as direct linguistic extraction, crisp rule conversion, and cluster-driven labeling enable transparent mapping and quantifiable interpretability metrics.
• Semantic enhancements through ontology-driven regularization and rule reduction techniques ensure compact, context-aware models tailored for practical applications.

A fuzzy rule-based model (FRBM) explanation is the systematic mapping of a fuzzy inference system—whose components include fuzzy sets, membership functions, and IF–THEN logical structures—into an interpretable semantic form for end users and auditors. The explanatory apparatus encompasses linguistic translation, structure-to-crisp conversion, constraint-driven interpretability enhancements, and model reduction techniques. Contemporary XAI research operationalizes these goals through diverse algorithmic and formal devices, spanning direct linguistic extraction, region-based crispifications, semantic regularizations, dynamic rule management, and cluster-preserving labeling, as detailed in multiple recent works.

1. Fundamental Definitions and Linguistic Mapping

A standard fuzzy rule $R$ in a neuro-fuzzy system for $n$ variables has the canonical form: $$R:\quad \mathrm{IF}~x_1~\mathrm{IS}~A_1~\mathrm{AND}\,\dots\,\mathrm{AND}~x_n~\mathrm{IS}~A_n~\mathrm{THEN}~y~\mathrm{IS}~B$$ Where $A_i$ and $B$ are fuzzy sets on their respective universes, parameterized by membership functions $\mu_A: X \rightarrow [0,1]$ (e.g., triangular, trapezoidal, Gaussian, singleton, sigmoidal, etc.).

Simiński & Wnuk present an automatic extraction method that translates any FRBM into a set of fluent English sentences by:

• Computing data statistics (mean $\bar x$ and standard deviation $\hat x$ for each variable)
• Reading each fuzzy set's numerical parameters
• Assigning linguistic labels to locations (e.g., "tiny," "medium," "giant") and degrees of fuzziness (e.g., "strictly," "loosely") using rule-defined intervals
• Combining these into atomic language phrases ("input 1 is moderately large")
• Composing full rule sentences ("If input 1 is loosely tiny and input 2 is loosely large, then output is strictly micro.")

Seven distinct location terms and five fuzziness terms are used, and the system handles overlapped fuzzy sets by mapping each to its dominant computed centroid descriptor. All rules are thus exposed as transparent, human-readable linguistic constructs fully derived from the FRBM's structure (Siminski et al., 2024).

2. Algorithmic Approaches to Extracting Explanations

2.1 Direct Linguistic Labeling

The extraction pipeline of (Siminski et al., 2024) formalizes linguistic mapping for all major fuzzy set types and rule forms (Mamdani, TSK). Importantly, this method:

• Handles both symmetric (e.g., triangular, Gaussian) and asymmetric (e.g., sigmoidal, arctan) descriptors
• Reduces multivariate antecedents and consequents to conjoined, narratively consistent English statements
• Achieves rule-level traceability for both simple and composite output structures

The process is robust to all standard fuzzy membership function families used in practice.

2.2 Crisp Rule Equivalent Extraction

The methodology of "Crisp complexity of fuzzy classifiers" (Fernandez-Peralta et al., 22 Apr 2025) transforms a fuzzy rule base into an exactly equivalent crisp rule system. This is achieved by:

• Representing each fuzzy set with a support hyperrectangle (e.g., for a trapezoidal $A = \langle a,b,c,d \rangle$)
• Mining all nonempty rule subsets $R\subseteq \mathcal{R}_{\text{fuzzy}}$ and their compatible regions $n$0 where specific rules uniquely dominate
• Generating, for each region, crisp IF-THEN rules of the form:

$n$1

• Calculating the normalized complexity $n$2 (where $n$3 is the original rule base size), thus providing a quantitative proxy for interpretability

This construction reveals the explicit partitioning induced by overlapping fuzzy sets and enables practitioners to reason about the combinatorial cost of explanations for arbitrary fuzzy classifiers.

2.3 Cluster-Driven and Attribute-Weighted Explanations

Interval type-2 fuzzy rule-based systems (Navarro et al., 2016) explain their decisions by associating each rule with a data cluster, defined via subtractive clustering. Each rule captures a well-separated region of the input space, characterized by a prototype $n$4 and associated interval-valued membership functions. Human validation proceeds by inspecting:

• Top-firing rules and their associated cluster centers
• Consequent weights reflecting class association certainties
• Aggregated contributions to the final decision

ADAR (Liu et al., 27 Apr 2025) extends rule and attribute-level interpretability by attaching learned, [0,1]-bounded importance weights to both features and rules. Attribute pruning (via thresholding and binary masking) and rule pruning/growing (via significance based on training error) yield concise models. For example, each resulting rule is of the form: "If Displacement is high around 120 (membership $n$5, attribute weight $n$6) and Weight is centered at 2800 ($n$7), then MPG is computed as $n$8Displacement$n$9Weight$": both the mechanistic structure and the feature importances are explicit (Liu et al., 27 Apr 2025).

3. Semantic and Ontology-Driven Explanation Enhancement

The ESC-Rules framework (Glauer et al., 2022) incorporates explicit semantic constraints through ontology-driven loss regularization:

• Rule-length penalties promote brevity
• Fuzziness penalties promote crisp, easily verifiable conditions
• Implication and mutual exclusion constraints reflect domain ontologies and user preferences (e.g., prohibiting logically inconsistent or redundant antecedents)

Fuzzy rules are weighted and composed in a linear structure, with degrees of participation for each literal determined by learned sigmoid-activated parameters. Rule extraction post-training involves thresholding to declare literal activity and reporting both the impact and "fit" of each rule, allowing domain experts to validate and directly interpret the system output.

4. Model Reduction, Rule Induction, and Complexity

The FRRI algorithm (Bollaert et al., 2024) generates concise fuzzy-rough rule bases through two-stage induction: rule pruning/generation and minimal covering optimization. Each candidate rule is built to guarantee no off-class coverage, is pruned to minimize length, and is selected as part of the minimal cover. Each rule is equipped with:

• Matching degree (degree to which input matches the antecedent)
• Confidence (derived from the fuzzy lower approximation)
• Coverage (degree to which the rule applies to each example)

In experiments, FRRI achieves high balanced accuracy (mean $$R:\quad \mathrm{IF}~x_1~\mathrm{IS}~A_1~\mathrm{AND}\,\dots\,\mathrm{AND}~x_n~\mathrm{IS}~A_n~\mathrm{THEN}~y~\mathrm{IS}~B$$0), compact rulesets (average $$R:\quad \mathrm{IF}~x_1~\mathrm{IS}~A_1~\mathrm{AND}\,\dots\,\mathrm{AND}~x_n~\mathrm{IS}~A_n~\mathrm{THEN}~y~\mathrm{IS}~B$$1 rules per fold, moderate rule length), and transparent conditions explicitly stating the type of similarity or dominance (e.g., "IF $$R:\quad \mathrm{IF}~x_1~\mathrm{IS}~A_1~\mathrm{AND}\,\dots\,\mathrm{AND}~x_n~\mathrm{IS}~A_n~\mathrm{THEN}~y~\mathrm{IS}~B$$2 AND $$R:\quad \mathrm{IF}~x_1~\mathrm{IS}~A_1~\mathrm{AND}\,\dots\,\mathrm{AND}~x_n~\mathrm{IS}~A_n~\mathrm{THEN}~y~\mathrm{IS}~B$$3 THEN $$R:\quad \mathrm{IF}~x_1~\mathrm{IS}~A_1~\mathrm{AND}\,\dots\,\mathrm{AND}~x_n~\mathrm{IS}~A_n~\mathrm{THEN}~y~\mathrm{IS}~B$$4") (Bollaert et al., 2024).

5. Advanced Explainability in Hierarchical and Additive Fuzzy Systems

Newer XAI frameworks extend fuzzy-rule explanations to high-dimensional, nonlinear, and reinforcement learning contexts:

• FAME ("Fuzzy Additive Model with Explainability") (Gokmen et al., 9 Apr 2025) integrates a three-layer pipeline: projection (dimension reduction), single-input fuzzy subnetworks (additive index model), and aggregation. Each subnetwork is constrained so that only two rules can fire for any given input (using "Gauss2MF" membership functions and non-crossing centers), yielding maximally concise, traceable explanations decomposable as the sum of local rule outputs. This construction mitigates rule explosion and enables per-dimension interpretability.
• In continuous-control DRL, fuzzy rule bases distilled from neural policies (via K-means clustering and TSK modeling) allow explicit explanation of agent actions. Performance and interpretability are jointly measured by metrics such as Fuzzy Rule Activation Density (FRAD), Fuzzy Set Coverage (FSC), and Action Space Granularity (ASG). For instance, a 4-rule triangular fuzzy surrogate in "Lunar Lander" achieves 97.83% action-matching fidelity, and extracted rules are readable as: "IF X is NEG (∼-0.61) AND Y is POS (∼1.98) ... THEN Action = [Main = 0.36, Side = 0.71]" (Araballi et al., 24 Feb 2026).

6. Impact, Metrics, and Limitations

Explainability of FRBMs is quantitatively assessed via:

• Linguistic fidelity (does the natural-language output capture rule semantics?) (Siminski et al., 2024)
• Normalized crisp complexity (proxy for the cognitive cost to understand the global partitioning, $$R:\quad \mathrm{IF}~x_1~\mathrm{IS}~A_1~\mathrm{AND}\,\dots\,\mathrm{AND}~x_n~\mathrm{IS}~A_n~\mathrm{THEN}~y~\mathrm{IS}~B$$5) (Fernandez-Peralta et al., 22 Apr 2025)
• Sparsity and overlap indices (e.g., $$R:\quad \mathrm{IF}~x_1~\mathrm{IS}~A_1~\mathrm{AND}\,\dots\,\mathrm{AND}~x_n~\mathrm{IS}~A_n~\mathrm{THEN}~y~\mathrm{IS}~B$$6 for rule overlap, $$R:\quad \mathrm{IF}~x_1~\mathrm{IS}~A_1~\mathrm{AND}\,\dots\,\mathrm{AND}~x_n~\mathrm{IS}~A_n~\mathrm{THEN}~y~\mathrm{IS}~B$$7 for fuzzy set spacing) (Liu et al., 27 Apr 2025)
• Rule activation focus and semantic constraints (e.g., FRAD, ontology loss) (Glauer et al., 2022, Araballi et al., 24 Feb 2026)

A plausible implication is that, while compactness and transparency are achievable, model reduction (as in (Bollaert et al., 2024) and (Liu et al., 27 Apr 2025)) must be balanced against the risk of reduced coverage. Similarly, converting fuzzy logic to crisp rules (as in (Fernandez-Peralta et al., 22 Apr 2025)) enhances non-fuzzy interpretability but increases rule count exponentially in the worst case.

7. Synthesis and Practical Recommendations

Modern FRBM explainability has converged on a range of complementary strategies:

• Direct extraction of human-interpretable language descriptions from fuzzy sets and rule antecedents/consequents is computationally tractable and robust to overlapping or ambiguous regions (Siminski et al., 2024)
• Structure-preserving rule reduction (via minimal covers, pruning, and dynamic adjustment) provides concise, mid-sized rule bases with explicit semantic weighting (Bollaert et al., 2024, Liu et al., 27 Apr 2025)
• Conversion to crisp rule partitions offers compatibility with non-fuzzy XAI paradigms and establishes a universal complexity metric for model selection (Fernandez-Peralta et al., 22 Apr 2025)
• Ontology-driven regularization and per-rule importance annotation furnish tools for domain adaptation and regulatory compliance (Glauer et al., 2022)
• Fuzzy-rule distillation now covers not only standard classification/regression but also advances into high-dimensional, time-dependent, and deep RL domains while preserving actionable, context-aware explanations (Gokmen et al., 9 Apr 2025, Araballi et al., 24 Feb 2026)

In sum, the explanatory apparatus for fuzzy rule-based models now supports both fine-grained, rule-level narrative explanations and global, complexity-driven audits, bridging XAI, cognitive transparency, and human verification (Siminski et al., 2024, Gokmen et al., 9 Apr 2025, Glauer et al., 2022, Fernandez-Peralta et al., 22 Apr 2025, Bollaert et al., 2024, Liu et al., 27 Apr 2025, Araballi et al., 24 Feb 2026).