| In the era of big data,high-speed,infinite,manifold data with concept drift is full of people ’s production and life.How to mine valuable information from massive manifold data has become a hot topic.Traditional data mining algorithms require processing batch arrival of independent and identically distributed data,which cannot be applied to dynamic data streams.It is urgent to study the classification of concept drift data streams,which has important value and practical significance.The classification task of concept drift data stream has the following two problems :(1)When a new concept appears,the model cannot be fully trained due to too few learning samples under the concept,which leads to low classification accuracy;(2)How to use the historical classifier to quickly restore the generalization ability on the new concept samples;(3)How to incrementally learn historical classifiers to improve generalization learning ability.In summary,considering the problems and research value of the concept drift data stream classification task,the research work of this paper is summarized as the following two aspects :For problem(1),a concept drift data stream classification algorithm LA-MS-CDC based on local classification accuracy is proposed.Firstly,k-means clustering and local classification accuracy are combined to select the optimal source domain classifier from the classifier pool.Secondly,the optimal source domain classifier and the target domain classifier are weighted and integrated to classify the samples.Thirdly,the loss of each classifier is calculated according to the real labels of the classification samples and the weights of the classifiers in the target domain and the source domain are updated.Then,the classification samples are used to update the target domain classifier and the optimal source domain classifier.Finally,update the classifier pool.For problem(2)(3),A concept drift data stream classification algorithm based on bidirectional learning Condor-CDC is proposed.After the algorithm detects concept drift,first,the reusable model is selected from the model library using local classification accuracy.Secondly,the bias regularization technique is used to add the reusable model to the regularization term of the PA classifier as a bias term to achieve model reuse.Finally,the reused model is used as the target domain classifier.The innovations of this paper are as follows :(1)The selection method of classifier and the maintenance method of classifier pool are proposed.According to the local similarity between source domain and target domain,the transferable source domain classifier is selected from the classifier pool.When the classifier pool is full,based on the characteristics of the component classifier,the diversity of the support vector is maximized to maintain the diversity of the classifier pool.(2)A model reuse technology is proposed.After detecting the concept drift,the reusable model is selected by using the local classification accuracy,and then the model is reused by combining the deviation regularization technology to alleviate the cold start problem encountered in creating the new model,thus improving the classification performance of the algorithm.At the same time,a dual weight update mechanism is proposed to selectively update the historical classifier,which improves the generalization performance of the historical classifier. |
