Use filtering techniques to improve the accuracy of association rules

Association rules' learning is a machine learning technique used to find interesting relations among data items and is a base to build an association rules classifier. The accuracy of the classifier highly depends on the quality and accuracy of data items. This accuracy can be affected negatively by noisy instances and this may lead to classification overfitting. This work investigates overcoming this problem by applying DROP3 or ALLKNN filtering algorithms to the datasets prior to generating association rules and building a classifier. Several experiments and comparisons were conducted to test the accuracy of the above filtering algorithms. The experiments were conducted on three noise levels: 0%, 5% and 10%. Results were more promising with ALLKNN as accuracy has improved remarkably especially with high noise ratios. With ALLKNN, average classification accuracy for the eight datasets in the 0% noise case improved from 70.47% to 73.63% compared to the base case when ALLKNN was not used. This improvement in classification accuracy was more apparent with the increase in noise ratio. In the 5% noise ratio the accuracy improved from 66.08% to 76.17%. In the 10% noise ratio the accuracy improved from 59.89% to 75.68%.