Hybrid neural systems for large scale credit risk assessment applications

摘要

Hybrid Neural Systems that integrate symbolic algorithms or fuzzy systems to Artificial Neural Networks (ANN) are a potential alternative to the more traditional ANN models. However, in contrast with the ANN models, these systems have not been yet fully explored from a practical viewpoint to show their effectiveness in large scale applications. This paper presents an extensive comparative analysis of the neuro-fuzzy models FWD (Feature-Weighted Detector) and FuNN (Fuzzy Neural Network), together with their rule extraction techniques in a large-scale problem. Two aspects are considered: generalization performance of the models, and the interpretation and explanation qualities of the extracted knowledge. The experiments are conducted in the context of a large scale credit risk assessment application in a real-world operation of a Brazilian financial institution. The results attained are compared to those observed with multi-layer perceptron networks.