The Semantic Classification Approach Base on Neural Networks

Yanli Shi,Jinxing Zhao

doi:10.1109/access.2020.2966227

Abstract

This paper addresses the classification problem, and a semantic classification approach using neural networks is proposed. The approach embeds the theoretical findings of the axiomatic fuzzy set theory in neural networks. Complex concepts are extracted by neural networks, which means that the class description is formed analytically rather than by tuning parameters of constraint conditions. The experiments are carried out on five benchmark datasets and compared results with five other neural network-based classifiers. The experimental results show that the proposed approach produces high classification accuracy and has a better explanation.

Highlights

Classification is frequently encountered task in many fields, such as economy [1], industrial engineering [2], natural environment [3]
Many approaches are applied to the classification problem, which includes fuzzy logical [4], support vector machine [5], artificial neural network [6]
The fuzzy sets and fuzzy logic are introduced to the classification problem, which makes the classifier has the ability of dealing with uncertain or imprecise data with high classification accuracy [7]–[11]

Summary

INTRODUCTION

Classification is frequently encountered task in many fields, such as economy [1], industrial engineering [2], natural environment [3]. The improvements in the proposed algorithm include the following aspects: 1) Take the membership degree of simple concepts as input, the algorithm fully considers the data structure and the semantic interpretation. Definition 3: For fuzzy concept ξ ∈ EM , μξ : X → [0, 1], {μξ |ξ ∈ EM } is called coherence membership function of the AFS fuzzy logic system (EM , ∨, ∧) and AFS structure (EM , τ, X ), if the following conditions are satisfied: 1) for ξ, ζ ∈ EM , if ξ ≤ ζ in lattice (EM , ∨, ∧), μξ (x) ≤ μζ (x); 2) if Aτ (x) = φ, μξ (x) = 0; 3) if Aτ (y) ⊆ Aτ (x), μξ (y) < μξ (x); if Aτ (y) = X , μξ (y) = 1. The details of the algorithm will be elaborated in the following

OBTAIN THE MEMBERSHIP OF EACH SAMPLE IN EACH

EXTRACT COMPLEX CONCEPTS BY USING BPNN

FORM THE CLASS DESCRIPTIONS FOR EACH CLASS

EXPERIMENTAL RESULTS

For each class do

CONCLUSION