Research On Imbalanced Classification And Its Coupling Relationship Based On Reinforcement Learning

With the improvement of information construction,various enterprise units and government agencies have accumulated lot of datasets.How to make reasonable use of scientific data mining methods to predict or classification data is the key to improving production capacity.Facing a large and complicated large amount of data,how to fully tap data value is the core of data work.Therefore,analysis and research data objects and data mining models corresponding to it are the focus of innovation and development of all walks of life.Imbalanced data exists in various real business scenarios,such as the stock price rising time sequence data,medical diagnosis data,and customer loss data.Due to the imbalance of data and the problems of multi-noise,and the overlapping data distribution,traditional machine learning models often have difficulty to achieve satisfactory results when facing imbalanced learning,or even unable to learn.On the other hand,the difficulty brought by data imbalances is that the errors of a few categories are divided into a small number of categories into a majority class.In real data,a few classes often have higher value.In order to solve the difficulty caused by data imbalance,the academic community has proposed a series of models for Imbalanced data.Because these algorithms have problems such as destructive data distribution,unable to adapt to data,and facing noise unstable,the effect is not satisfactory.This article first analyzes the interactive effect of the model error distribution in the classification boundary.Based on this analysis,the dynamic classifier system proposes a soft resampling dynamic classifier ensemble selection algorithm,and the algorithm is promoted from resampling to reweight.Finally,this article consolidates the sampling parameter as a meta parameter from the model,builds a meta-sampler,and trains the meta sampler with reinforcement learning algorithm to realize the dual adaptive learning of the model and data.There are mainly the following work in this article:This paper uses data subsets with different error distribution to train multiple homogeneous classifiers,and draws the decision-making boundary visual diagram of each classifier to analyze the interactive mechanism of the model error distribution and the classification boundary.By analyzing,it can be found that the classification boundary will be offset with different datasets,and the base classifiers obtained by using hard samples have clearer and delicate decision-making boundaries.On the whole,the base classifier based on error distribution training presents regional and diversity characteristics.This article proposes a Soft Resampling Dynamic Ensemble Selection model on the basis of the analysis of the base classifier.The model contains two algorithms,the base classifier generates algorithm and dynamic prediction algorithm.Base classifier generation algorithm takes into account the deviation and variance advantages of static ensemble learning,while.At the same time,it has diversity and regional characteristics.The dynamic prediction algorithm proposed in this article is based on the classifier evaluation function,and the performance of the performance of the classification device in the field of the prediction point is given the prediction weight,which enhances the local information mining ability of the multi-classifier system.The confidence in the prediction point is incorporated into the evaluation function inspection factors,which improves the rationality of the evaluation mechanism.This article conducts parameter sensitivity tests on Soft Resampling Dynamic Ensemble Selection model,and the results show that Soft Resampling Dynamic Ensemble Selection model is rubout for parameters within a specific range.For this result,this article gives the advice of parameters selection of Soft Resampling Dynamic Ensemble Selection model.This article compares Soft Resampling Dynamic Ensemble model with the representative algorithm field of dynamic classifier systems and imblanced ensemble learning fields on the real dataset.The experimental results show that the Soft Resampling Dynamic Ensemble Selection model has the optimal performance.in stability and accuracy.In order to expand the application scenarios of the Soft Resampling Dynamic Ensemble Selection model,this article will promote the classifiers generate algorithm of the Soft Resampling Dynamic Ensemble Selection model,and propose a Soft Reweight Dynamic Ensemble Selection model.Unlike Soft Resampling Dynamic Ensemble Selection model,Soft Reweight Dynamic Ensemble Selection model will promote most class sample probability to most sample weights,and design cost sensitive factors to balance class deviation.In this paper,the Soft Reweight Dynamic Ensemble Selection model and the Soft Resampling Dynamic Ensemble Selection model are tested on the real datasets.The experimental results show that the classification performance of the two models are close,and the performance of the Soft Reweight Dynamic Ensemble Selection model has a weak advantage due to the increase of training samples.In this paper,the sampling parameter of the Soft Resampling Dynamic Ensemble Selection model are decoupled and set as the meta parameter.The training meta sampler guides the learning of the ensemble framework.Based on the understanding of the error distribution of the model,this paper designs a meta state containing training process,error distribution and other information,and shapes a reward function to train the meta sampler with the reinforcement learning algorithm.The meta sampler and the Ensemble framework constitute the Meta Soft Resampling Dynamic Ensemble Selection model.The model realizes dual adaptive learning of data and model,and has the ability of cross task migration.In this paper,the performance of Meta Soft Resampling Dynamic Ensemble Selection model is tested on real data sets.The experimental results show that the classification performance of Meta Soft Resampling Dynamic Ensemble Selection model is close to Soft Resampling Dynamic Ensemble Selection model,and has been improved to a certain extent.Finally,this paper conducts a cross-task transfer experiment on the Meta Soft Resampling Dynamic Ensemble Selection model,and the experimental results show that the meta-learner has the ability of cross task transfer.