Optimized random forest unbalanced data set processing method

Abstract

The invention discloses an optimized random forest unbalanced data set processing method, which comprises data preprocessing, random forest model construction and classification prediction, and is characterized in that k majority class samples nearest to minority class samples are found out in the data preprocessing part to form areas difficult to distinguish. the samples of the region are re-tagged in the original data set, minority class samples are generated in the region which is difficult to distinguish, and the re-tagged original data and the region which is difficult to distinguish is outputted after the samples are newly added as different training sets; the two data sets processed by the data preprocessing part are used as training sets of the model in the construction of the random forest models to obtain two random forest models, the classification prediction enters the two random forest models for verification in two stages, and finally a classification prediction result of a sample is obtained. The purposes that the minority class prediction performance is improved, and meanwhile the majority class prediction accuracy cannot be seriously reduced are achieved.

Description

technical field [0001] The invention belongs to the technical field of data analysis, mining and machine learning, and in particular relates to an optimized random forest method for processing unbalanced data sets. [0002] technical background [0003] With the advent of the era of big data, data mining has become an increasingly important technology, and classification is the most common task in data mining. Using classification algorithms to mine the potential information of data is conducive to providing effective predictions for problems. In real-world classification scenarios, there are often situations where many data sets are unevenly distributed, and for different problems, different classifications have different degrees of emphasis. The general classification algorithm seeks to improve the overall classification accuracy of the data set, resulting in the prediction classification accuracy rate of the minority class samples being much lower than the prediction class...

AI Technical Summary

Problems solved by technology

[0010] The disadvantages at the data processing level are: oversampling technology directly generates similar minority samples because it does not analyze minority samples, which may easily lead to redundant samples and lead to model overfitting

Undersampling technology reduces the majority class samples to reduce the imbalance rate of the data set, resulting in the loss of most class information and reducing the classification accuracy of the majority class.

[0011] The disadvantage of the ENN algorithm is that even if the algorithm removes some samples of the majority class, the distribution of the data set may still have a large imbalance rate, and because some samples of the majority class are deleted, the classification accuracy of the majority class will be reduced. decline

[0012] Although the biased random forest algorithm with the best effect at present achieves the purpose of improving classification performance by finding error-prone areas and training random forests through two data sets, it throws less minority class information and obtains the first The two data sets may still be unevenly distributed, and because Random Forest uses Bootstrap random resampling technology, this will reduce the probability of minority class samples being sampled and affect the classification accuracy of minority class samples

Images