Extracting Positive and Negative Association Classification Rules from RBF Kernel

Recently, building associative classifiers by miming association rules is a hot research problem. As negative association rules also help to understand the data, in this paper, we present our InterRBF algorithm, which expands RBF kernel into its Maclaurin series, and then mines positive and negative association rules which make great contribution to classification from this series, so as to learn association classifier from the SVM classification model. Taking {0,1}ⁿ as input space, we also show the reasonable value field of hyper-parameter g of RBF kernel by applying the theory of Occam´s razor, so as to have good classification performance. Experiment results on 6 UCI datasets show that InterRBF could build associative classifiers with better accuracy and smaller size of rule set than ARC-PAN[1], another associative classifier which is also build with both positive and negative association rules. Furthermore, compared with CMAR [3] and CPAR [4], the average accuracy of InterRBF over the 6 datasets also outperforms the two classifiers.