Mining Top-K Classification Rules

Abstract

In this thesis we present a classifier that uses the associative classification approach. We exploit the mined top-k pattern, to extract classification rules from a set of data to performs a classification based on predictive association rules. The pattern top-k extracted by the set of data are approximated pattern that are able to briefly describe the input data. The top-k pattern discovery problem is commonly stated as an optimization one, where the goal is to minimize a given cost function, watching the accuracy of the data description. These patterns extracted, are used as classification rules in prediction, within our rule-based classifier. For generating of candidate rules, we used a different approach, which consists in adopting a greedy algorithmic framework named PaNDa+ to generate rules directly from training data. Once we extracted the rules, we carried out a pruning, and calculated the prediction power of each rules, to obtain the best rules in prediction. We evaluated the goodness of the classifier by measuring the quality and the accuracy of the extracted rules. The evaluation was conducted on synthetic data sets, and the results compared with other classifiers as JCBA, CPAR, Weighted Classifier, SVM, C4.5.