Handwritten Digit Recognition Method Research Based On Belief Decision Trees

With the rapid development of science and technology,more and more information strikes into human production and life,a large scale of which is with different kinds of uncertainty,which makes it a great challenge in the development of information and technology to obtain effective information from uncertainties.For example,as a popular research topic,handwritten digit recognition has always been widely concerned.Although researchers can get large amount of data as training set samples in this issue,most of the data is with more or less epistemic uncertainty due to the existence of imprecise,unreliable and so on.Usually in engineering practice,researchers deal with the problem by labeling uncertain data manually to get precise training data.While both the selection of samples to be marked and the addition of labels require manual participation,which makes the labor cost increase when more data is gotten.How to effectively deal and even build classifiers with uncertain information becomes a problem that a number of researchers need to solve extremely.For its ability of modelling and reasoning with kinds of uncertainties,belief function theory has been widely researched and applied in to many fields such as medicine and engineering etc.since proposed.Separated with traditional information fusion and evidence reasoning,the research of belief function has stepped onto a new stage thanks to the return of evidential statistical inference recent years especially after 2008.On this basis,a few researchers take the lead in combining belief function theory with machine learning,breaking research gaps on this field,and reach to good results.Moving on with previous work,we combine belief classification tree with ensemble algorithm bagging to get the final classifier by integrating many single and simple belief trees together on the basis of modelling cpistemic uncertainty by mass function.In this work,the belief classification trees used as base classifiers are learned with uncertain training samples directly.In the meantime,considering the fact that most uncertain classification methods sidestep on practical application,to verify the peerformance of ABC4.5 and BGBC4.5 algorithms,we apply these two methods into handwritten digit recognition problem respectively and get good accuracy.Classifiers are modeled and built based on the training sets whose output part exists a large number of different kinds of epistemic uncertainties in both experiments.We analyze the advantages and superiorities of BGBC4.5 algorithm by comparing with the diverse performances of other methods when facing varying degrees of uncertainties and data quality.