Learning Based K-Dependence Bayesian Classifiers

In this paper, we introduce a new mining algorithm to improve the classification accuracy rates aiming at the deficiency of the typical K-Dependence Bayes (KDB) model which ignores the topology changing as the result of inputing of test instances. Under this condition, we put forward an algorithm called Base K-Dependence Bayes (B-KDB) which consists of a Label-based K-Dependence Bayes (L-KDB) algorithm and a Instance-based K-Dependence Bayes (I-KDB) algorithm. The I-KDB algorithm is used to build I-KDB model by instances to be tested and it can deal with the problem of test instances topology keep on changing. However I-KDB model is extraordinarily sensitive to data and it may suffer from overfitting and the effect of noisy instances, therefore L-KDB algorithm is designed as complement. After combing these two algorithms into B-KDB, we built B-KDB model and tested the performance against the KDB model in classification accuracy, precision, sensitivity-specificity analysis with 10-fold cross validation on 55 real benchmark datasets from University of California Irvine (UCI) machine learning repository. The experimental result, which shows the classification accuracy of our model twice as much as KDB, indicates our algorithm efficient and proves our idea of improving the KDB algorithm classification accuracy feasible.