A rule-based classification approach called Associative Classification (AC) normally constructs accurate classifiers from supervised learning data sets in data mining. It extracts "If-Then" rules, and associates with two computed parameters each of the generated rules; support and confidence. These two parameters are used to determine the dominance of the laws during the creation of a classification. In current AC algorithms, all of its corresponding training data is discarded whenever a rule is inserted into a classifier. The discarded data, however, are actually used to calculate support and confidence for other rules and will affect other lower-ranked rules, as rules normally have common examples of training data. Use static support and confidence will result in very large, less accurate classifiers. Therefore, it is necessary to provide a method that modifies the support and confidence of other rules. This paper proposes a new procedure called Active Pruning Rules (APR) to overcome the above problem, in order to further improve the performance of the classifiers-especially predictive accuracy and reduction of rule redundancy. The experimental results obtained from a variety of data sets from the University of California Irvine (UCI) and real data set for adult autism classification showed that APR is highly competitive with other AC and rule-based classifiers and also produces smaller and more accurate classifiers.
Results:
cd APR
set dataset path in validation.py
python validation.py