Unbalanced data classification method

Abstract

The invention relates to an unbalanced data classification method. The unbalanced data classification method comprises the following steps: in a labeled data set L, firstly processing each data point, calculating the distances between the data points and all non-similar data points, and reserving the shortest distance as the characteristic of the data points; arranging all the data points according to the characteristic from small to large, taking the first t data points with the minimum characteristic to form an initial training set T, and enabling the remaining data points to form an initial non-training set N; using a support vector machine, and utilizing an active learning strategy to carry out iterative learning on the training set T; after training begins, a temporary classification hyperplane P is generated in each step of iteration, using the Ps to carry out trial classification on all the data points in the N, if mispredicted data points exist, drawing an item at random from the mispredicted data points to be added to the training set T, and meanwhile selecting the data point closest to the P in the N to be added to the T; if no mispredicted data points exist in the N, selecting the data point closest to the P from the data points to be added to the training train T. carrying out subsequent training.

Description

technical field [0001] The invention relates to a method for classifying unbalanced data, which belongs to the field of computer data analysis and mining, in particular to a data classification algorithm. Background technique [0002] Imbalanced datasets, that is, datasets that have large differences between the number of samples each class has. In the binary classification of unbalanced data sets, the class with a small number of samples is usually called a positive class, and correspondingly, the class with a large number of samples is called a negative class. Data imbalance is very common in current applications, such as medical diagnosis, intrusion detection, fraud prevention, distinguishing things from satellite images, etc. When classifying, the classification accuracy of the positive class is our main concern. For example, in disease diagnosis, the misdiagnosis of healthy people will be resolved during reexamination, but the misjudgment of cancer patients as normal ...

AI Technical Summary

Problems solved by technology

However, at the beginning of the training, the support vector machine of active learning determines the initial training set by random extraction from all data points. The randomness is very large, and it is commonly used to select the data point closest to the current classification hyperplane. The method of adding the training set to adjust the classification hyperplane, the adjustment of the hyperplane is limited, and some final irrelevant data will still be introduced in the iterative training process, these problems will adversely affect the final classification effect

Images