Fuzzy Case-Based System for Classification Tasks on Missing and Noisy Data

Hybrid models combine different technologies to obtain a product that shares their advantages and minimizes their deficiencies. The solutions given by a case-based system (CBS) rely on similar past experiences, which are commonly described in terms of both symbolic and continuous attributes. The nearest neighbor (NN) principle commonly followed to develop CBS for classification task proceeds from the assumption that similar cases have similar solutions, having the definition of the distance (similarity) function a central attention for obtaining a good accuracy on a given data set. This paper presents a hybrid model to solve classification tasks using NN principle but including generalized knowledge from the set of given instances to improve the performance in contrast to pure lazy learning algorithms. The fuzzy case-based system, referred as FuCiuS, interprets predictive numeric attributes in terms of fuzzy sets defining in a conceptually uniform way a one-dimensional (local) criterion to compare mixed and missing data. Experimental analysis show good performance for FuCiuS in comparison with well-known classifiers on missing and noisy data, while a more natural framework to include expert knowledge by using linguistic is provided guaranteeing both robustness and interpretable solutions.