Associative Classification with Knn

Associative classification usually generates a large set of rules. Therefore, it is inevitable that an instance matches several rules which classes are conflicted. In this paper, a new framework called Associative Classification with KNN (AC-KNN) is proposed, which uses an improved KNN algorithm to address rule conflicts. Traditional K-Nearest Neighbor (KNN) is low efficient due to its calculation of the similarity between the test instance and each training instance. Furthermore, the accuracy of KNN is largely depended on the selection of a “good value” for K. AC-KNN generates for each test instance a specific training set composed of instances covered by best n rules which match the test instance. Thus, the nearest neighbors from the specific training set are not only similar to but also associative with the test instance. As a result, such nearest neighbors will make better decision on classifying a conflict instance. Our experiments on 12 UCI datasets show that AC-KNN outperforms both AC and KNN on accuracy. Compare with KNN, ACKNN is more efficient and more stable to the number of nearest neighbors.