Research On Confidence Machine Learning Methods Based On Controllability

Study on machine learning algorithms has achieved great progress,and machine learning algorithms have been applied widely;however,most machine learning algorithm can only give such simple judgments as "yes" or "no","belonging to" or "not belonging to",lacking a confidence mechanism to explain how much the credibility or reliability of such judgment is.A confidence machine is such a machine learning method with confidence mechanism.A confidence machine is further extension of many machine learning methods,and it can not only make performance prediction like many machine learning methods but also measure the quantitative quality in the prediction results to give credibility and confidence.The current research efforts on confidence machine learning algorithm typically have profound theoretical basis,complex algorithm,very few functions for alternative calculation,and uneasy to understand and use.This dissertation mainly studies machine learning methods based on controllable confidence,to find a simpler,more efficient,more reliable and more practical machine learning method with controllable confidence,and the main study content is as follows:Such questions as presentation of confidence machine question,confidence,were discussed first.The confidence mechanisms were divided into the confidence mechanism of overall average confidence learning method,the confidence mechanism of Bayesian learning method,the confidence mechanism of transductive learning method,and the confidence mechanism of learning method with reject option.All research activities in this dissertation were realized based on the fourth confidence mechanism,i.e.,the confidence mechanism of learning method with reject option.For the question about two-class confidence classification,an algorithm,i.e.,two-class confidence classification based on one class classifier(TCCC-OCC)algorithm,was proposed.This algorithm realized delimiting of acceptance region and rejection region through calculation and analysis of the results from two times of learning of identified samples,thus to omit the calculation of specific.confidence necessary for each unknown example in traditional confidence machine learning and the setting of threshold of rejection region,reducing the calculation workload.In addition,multi-layer learning was conducted for the learning results using the integrated learning method,to further raise the recognition rate.The question about confidence of controllable rate was studied,and a controllable confidence classification based on two-class classifier(CCC-TCC)algorithm was proposed.This algorithm learned samples using a support vector machine(SVM),and then converted the learning results from spatial values to one-dimensional output score values.The order of magnitude of the output score values from the SVM kept the order of distance from the hyperplane in SVM classification,so the confidence and error rate could be controlled by setting the threshold.The CCC-TCC algorithm included four sub-algorithms,i.e.3 controllable confidence classification sub-algorithm with total error rate of classification being set,controllable confidence classification sub-algorithm with error rates of classification of positive and negative examples being set respectively,controllable confidence classification sub-algorithm with percentage of output conversion value being set from score,and controllable confidence classification sub-algorithm with percentage of output conversion value being set from score of classification error,and they were demonstrated through experiment on 5 datasets,including heart disease,and diabetes mellitus.The controllability of confidence regression question was studied,and an algorithm of confidence regression based on k-nearest neighbor(CR-KNN)was proposed.This algorithm realized delimiting of acceptance region and rejection region by conducting error judgment of regression learning results with the KNN algorithm as a tool,thus to realize confidence regression,and realized the control of confidence regression by setting specific change in error value.