Study On Hybrid Classifier Based On Bayesian Network And BP Neural Network

With the coming of the era of big data, the demand on classification for massivedata is increasing, and various classification algorithms have been applied in differentfields. However, In the face of rapid growth of the data in information age, singleclassification algorithm usually has drawbacks, which unable to meet the differentdemands. Experimental results have demonstrated that the hybrid classifier, which isconstructed by different algorithms, can effectively utilize the different advantages ofthealgorithm, thereby exhibitingbetterclassification performance.BP neural network is the most widely used and rapidly developed algorithm in thearea of artificial neural network. However, BP neural network classifier has a slowlyconvergence rate and also can easily fall into local optimum. In order to improve theperformance of this classifiers, there is an improvement on BP algorithm which isaiming at the slowly convergence rate. This paper proposed a thought that integratingthe conditional log likelihood (CLL) into the supervised BP neural networkclassification. In the process of classification, calculating the conditional probability oftest samples by making full use of the advantage of CLL, and also using the conditionalprobability to increase or reduce the corresponding weights in the error backpropagation, which can simplify the calculation in the error feedback process. The BPalgorithm is improved according to the above thought, and the experimental tests for themodified algorithm have made on convergence speed and accuracy, which illustrates theeffectiveness and practicality of this algorithm. Yet, when the number of attributes isbeyond a certain scale, the improved BP neural network classifier still exist the problemthat can easily fall into local optimum. However, the Bayesian network classifier thatmaking use of the conditional independence test in the process of structure learning canconstruct an efficient network structure which is highly fitted with the data. Thealgorithm can obtain global optimal solution under some certain conditions, so that theimprovement can help to avoid the local optimum effectively. Bayesian networklearning has been proved to be a NPhard problem, but it also has its obvious advantageswhen thestructure is relatively simple.In consideration of the complementary advantages between the improved BPneural network algorithm and the Bayesian network algorithm based on conditional independence test, this paper mixing these two algorithms by using the thought ofsurvival of the fittest. The hybrid algorithm always selects the classifier which has ahigher fitness with the data sets as the final classifier. The experiment mainly collecteddata on the follow two indicators, convergence speed and accuracy; it has proved thattheproposed hybrid classifier can obtain good classification performance.