Searching extended emerging patterns for supervised classification
Milton García Borroto · LA Referencia (Red Federada de Repositorios Institucionales de Publicaciones Científicas) · 2010
For many learning tasks, a high accuracy is not the only desired characteristic of a supervised classifier; A classifier should also be easily comprehensible by humans. Although higher classification accuracies are usually obtained at the expense of classification comprehensibility, Emerging Pattern classifiers are both accurate and easy to understand. The main contribution of this dissertation is the introduction of two new kinds of emerging patterns, which are more expressive than traditional definitions: Extended Crisp Emerging Patterns and Fuzzy Emerging Patterns. The higher expressiveness of the new patterns allows to obtain more accurate classifiers, without sacrificing understandability. Another contribution of this dissertation is a collection of algorithms for mining the new kinds of patterns from a database containing mixed and incomplete data. The classifiers proposed in this dissertation, using the new patterns, attain higher accuracy than traditional emerging pattern classifiers and other comprehensible classifiers, while they are competitive with state-of-the-art non-comprehensible classifiers. The selection of the classifier to be used in a particular problem depends on the type of patterns (crisp or fuzzy) the user wants to obtain, and a tradeoff among accuracy, complexity, and classification speed.