Research On Method Of Novel Class Detection And Classification For Concept-Drifting Data Stream Mining

In recent years,data stream mining technology has attracted many scholars' attention.But,the traditional data stream mining technology still has serval urgent problem in concept drifting and concept evolution in different application field.Therefore,this paper presents an improved novel class detection and classification algorithm for concept drift data mining,and focuses on the classification algorithm base on data steam feature space conversion method.The primary research works of the thesis are given as follows:1.We classified and summarized a part of common the concept drifting data stream classification algorithms algorithms between 2000 and 2016.And we designed experiments betweent different categories of algorithms,and summarized the performance difference between different categories of data stream classification algorithms.We pointed out some problems existing in the concept drifting data stream classification algorithms.And we gived reasonable improvement directions.2.We proposed Classfication and Novel Class Detection Algorithm Base on Mahalanobis Distance for Concept Drifting Data Stream(C&NCBM).This algorithm can be used as a similarity measure by introducing markov distance instead of Euclidean distance.It uses Mahalanobis as similarity measure.And it taking into account the correlation between the sample's attributes and the effect of the small variation of the variables,can effectively detect and mark the novel class(es)that appear(s)in the concept drifting data stream,and update the classification model to adapt to changes in data stream to improve the accuracy of the algorithm.In the artificial data set and UCI data sets,we compared the algorithm's classification performance and ability to handle concept drifting.Experimental results show that improved Novel Class Detection Algorithm Base on Mahalanobis Cohesiveness and Separability is effective and feasible,and it improved the classification accuracy and reduced classification evaluation time and it can handle the concept drifting of data stream.3.We proposed a method named Algorithm of Local Lossless Homogenizing Conversion Base on ReliefF(LLHCCR),it sets the threshold value for the feature space.If the threshold value is exceeded,use ReliefF to filter the feature attributes,otherwise use lossless homogenizing conversion.Our method reduces the possibility of the dimensional disaster of the existing lossless feature space transformation dimension,keeps the classification accuracy of the algorithm,and significantly reduces the evaluation time of the algorithm.In the artificial data set and UCI data sets,we compared the algorithm's classification performance and ability to handle concept drifting.Experimental results show that the Proposed Local Lossless Homogenizing Conversion of data stream feature space based on ReliefF is effective and feasible,and it improved the classification accuracy and reduced classification evaluation time,and it also has the ability to deal with the concept drifting of data stream..The research contributions of this paper are as follows.Summarized a part of common the concept drifting data stream classification algorithms algorithms between 2000 and 2016.The traditional data flow classification algorithm is improved from the aspects of feature space conversion and novel class detection.