Research On Fast Classification Algorithm Of Multiclass Relevance Vector Machine Based On Clustering

With the continuously improving of computer performance and rapid growth of Internet,all kinds of information appeared a growth of exponential trend. It become an important andurgent research topics to using technologies, including statistical methods, machine learning,data mining to extract valuable information and handling the practical problems of productionand life. As an important technical method Classification has broad application prospects. At2001, for the first time Michael E. Tipping proposed the concept of relevance vector machinebased on sparse Bayesian learning theory. This kind of learning machine has advantages, suchas sparsity model and probability output. But due to its high time complexity in the trainingprocess, the scale of RVM training data set is very limited. For those issues above, in thisarticle the following relative researches and analyses are made.For the first part, based on the study of the current state-of-art of RVM ClassificationAlgorithm at home and abroad, this paper proposes a fast classificing algorithm: BF-mRVM.This algorithm utilizes local learning strategies and introduces the Bit Reduction Algorithm tocluster Large-scale training sample, then divides the training data into scaled clusters of data.Then, for the cluters that have more than one type of data, mRVM classifier is utilized toconduct the local training.For the second part, in this article the test sample set takes the same cluster method astraining sample set, to devide the test sample set into a plurality of test sample cluster, thencalculate the haiming distance of the binary code between test sample cluster and all thetraining sample cluters one by one. Then select the nearest classifier, which has the nearestdistance to conduct testing. This is a dynamic classifier selection process.For the third part, through8groups of Data sets experiments, this article indicate thatBF-mRVM Algorithm can not only the improve training speed, but also can obtain a sparsityof model to some extent, most important the ability of dealing with large-scale datasets.Finally, this article discusses the effect between barameter b in the Bit Reduction clusteralgorithm and model training time, the number of vectors, test accuracy and testing time.Through analysis, conclusion can be made that when parameter b increased gradually, thisfour measures gradually tend to smooth.