Imbalanced Data Classification Analysis Based On Generative Adversarial Networks And Reinforcement Learning

In the application scenarios that need to achieve classification tasks,the widespread imbalance of category sample data sets is a difficulty faced by machine learning algorithms.With the rapid development of intelligent computing,Internet and other fields,unbalanced data sets with larger scale and higher imbalance rate are constantly generated.In order to deal with the classification difficulties caused by unbalanced data sets,this thesis will propose an algorithm framework combining generative adversarial networks and reinforcement learning from two aspects of data generation and classification algorithms,which can alleviate the problem of unbalanced data while improving the ability to identify the minority samples.Among them,the generative adversarial network is applied in the data resampling stage,and new in-distribution data samples are generated through the adversarial learning of the discriminator and the generator,increasing the training opportunities of the agent.Use the reward mechanism in reinforcement learning to train agents to achieve classification tasks,and emphasize the recognition of minority samples through the difference in the size of rewards for different types of samples.The algorithm framework proposed in the thesis has the following three advantages for dealing with the problem of unbalanced datasets:(1)By emphasizing the minority samples through the reward mechanism,it can improve the agent’s ability to classify minority samples.(2)By generating new data generated by the adversarial network,the sample expansion based on the original data distribution is realized,which can be used to balance the size of the reward and alleviate the problems caused by the imbalance of the agent training data.(3)The data set generated by learning the data distribution can increase the chance of agent training and improve its ability to explore data.In the reinforcement learning algorithm,each sample composed of features is used as a state,and the reward of the environment is combined with the correctness of the agent’s prediction of the sample label.Based on this,the Markov decision process is established,and the supervised classification task is transformed into a sequential decision task.The generated data and the original data are randomly mixed to generate a new data set as the training data of the agent.Through continuous trial and error and reward feedback,the agent is guided to realize the classification task of different types of samples.Through experiments on real data sets,it shows the feasibility of applying reinforcement learning to train classifiers and the effectiveness of using generative adversarial networks to generate new data to improve classification capabilities.Through the experiment on the unbalanced data set and comparing with the classification results of other machine learning algorithms,the recall rate,F1 value,Matthews coefficient and other evaluation indicators show that reinforcement learning has better recognition ability for minority samples.Generative adversarial network has improved the classification ability of agents.