| In recent ten years,the unbalanced data classification algorithm has become one of the important research directions in the field of machine learning.In the problem of unbalanced data classification,due to the extremely unbalanced sample data between categories,a few class samples are easily swallowed by the majority of class samples,which makes the classifier model unable to effectively identify a few classes when classifying.In real life,the application of unbalanced data classification is extensive,such as network intrusion detection,bank fraud monitoring,machine fault diagnosis,etc.Therefore,it is important to study how to improve the classification algorithm for the classification performance of unbalanced data.This paper studies and analyzes the problem of unbalanced data classification.On the basis of mastering the latest unbalanced data classification methods,it starts from the under-sampling and over-sampling methods in the data level,combined with integrated learning at the same time,a targeted solution to practical problems is proposed to improve the application value of the unbalanced data classification algorithm model.The specific work is as follows:(1)Aiming at the problem that the traditional classifier model is difficult to accurately predict the extreme temperature in the temperature class,a clustering undersampling integrated classification algorithm based on density peak is proposed.Firstly,the Pearson coefficient correlation method is used to select the temperature prediction factor,and according to meteorological knowledge,the Jiangsu regionalized temperature classification table is established at the same time;Secondly,the cluster centers generated by the clustering of the density peak algorithm on the majority of samples are undersampled according to the sampling rate to reduce the imbalance of the data;Finally,the balanced data is brought into the self-adaptive enhancement model based on BP neural network classifier,and the weight distribution of samples is changed to improve the classification performance of the algorithm model for unbalanced data.The experimental verification is carried out on the selected historical daily meteorological data sets of seven stations,and the proposed algorithm can effectively identify the extreme temperature level compared with the other three traditional algorithms.(2)Aiming at the problem that oversampling method in software defect prediction exists samples overlap and the low misclassification value of synthetic samples,a clustering oversampling ensemble classification algorithm based on distance weight is proposed.Firstly,based on the idea of bottom-up aggregation strategy of AGNES algorithm,when hierarchical clustering is carried out for each minority class subcluster,the existence of majority class samples is screened,and then multiple minority class subclusters without majority class samples are generated;Secondly,weights are assigned to each minority sample according to the average Euclidean distance of each sample in each minority subcluster and its K nearest neighbor majority samples,and the minority samples are oversampled according to the weight distribution;Finally,the clustering oversampling method based on distance weight proposed in this paper and ROS,SMOTE are used to construct a random forest model,giving full play to the overall advantages of the integrated model.Experimental comparison results based on NASA data sets show that the proposed algorithm model can effectively improve the prediction performance of software defect modules. |
