Obtención de reglas de clasificación difusas utilizando técnicas de optimización - caso de estudio riesgo crediticio

Identificador:  TDX:3116

Handle: https://hdl.handle.net/20.500.11797/TDX3116

Autors:  Jimbo Santana, Patricia Rosalia

Resum:
The central contribution of this thesis is the definition of a new method designated as FRvarPSO (Fuzzy Rules variable Particle Swarm Oprmization) that generates a pool of fuzzy classification rules of easy to understand, low cardinality and good accuracy. These characteristics aid at identifying the relationships in data facilitating and justifying the decision making process. FRvarPSO combines a competitive neuronal network, with an optimization technique based on clusters of particles with variable population for the obtention of fuzzy classification rules, able to operate on nominal and numeric attributes. The antecedents of the rules are formed by nominal attributes and/or fuzzy conditions. The formation of conditions requires to know the membership degree of the linguistic variable to the different fuzzy subsets. This thesis proposes three distinct alternatives for the obtention of the fuzzy subsets, and their membership function. One of these options was partitioning the rank of each numeric attribute at intervals of equal length, and centering each of them in a triangular function with a suitable overlapping. Another of the forms to obtain the fuzzy subsets has been utilising Fuzzy C-Means method. Additionally, we used also as a technique based on the knowledge of an expert. FRvarPSO is applied to different databases of the UCI repository and to a particular case of study of credit scoring. In this last case, we information of borrowers and macroeconomic contextual information, obtained from real databases of financial institutions of the Ecuador. This method contributes ken at the computer area and has evidenced that also realises contributions at the area of the economics.