A carregar...
Publicação
UnimNeuron
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 489.57 KB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
In evolving data streams, neuro-fuzzy systems must adapt to non-stationary distributions while preserving interpretability and numerical stability. However, most evolving fuzzy models rely on fixed logical connectives or implicitly encode logical behavior, limiting transparency and flexibility. This paper introduces the UnimNeuron, a neuro-fuzzy neuron whose antecedent connective is automatically selected from data using an order-invariant unimum-based aggregation. The proposed neuron dynamically exhibits conjunctive (AND), disjunctive (OR), or compensatory (COMP) behavior without operator search or manual tuning, and provides an explicit index to inspect the logical regime of each rule. Based on this formulation, two evolving neuro-fuzzy classifiers are proposed and evaluated under a prequential protocol on real and synthetic data streams with concept drift. Experimental results show that the proposed models achieve predictive performance statistically comparable to state-of-the-art evolving fuzzy systems, while maintaining a compact rule base, zero numerical instabilities, and interpretable fuzzy rules. An ablation study further confirms the relevance of automatic connective selection across diverse scenarios.
Descrição
de Campos Souza, P. V. (2026). UnimNeuron: Automatic Connective Selection for Interpretable Neuro-Fuzzy Rules. Paper presented at 2026 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE), Maastricht, Netherlands. https://linklings.s3.amazonaws.com/organizations/WCCI/wcci2026/submissions/stype108/SLRxV-fuzz_pap124s2.pdf
Palavras-chave
Fuzzy logic neuro-fuzzy systems uninorms interpretability aggregation operators