The Research On Deep Reinforcement Learning For Imbalance Classification

One main challenge for machine learning research is to learn a classifier from imbalanced dataset,since they will be biased towards the features of the majority group,which makes it hard to identify the minority group.At the same time usually the minority class is more important,as despite its rareness,it may carry more important and useful knowledge.The conventional classification algorithms for imbalanced data may not work when the data distribution is extremely imbalanced,and they cannot be applied to a general scenario.To address this,in this paper we propose a Deep Reinforcement Learning based method by designing a universal model for more imbalanced classification problems.This paper is organized as follows:(1)From a theoretical point of view,we first analyze the feasibility of utilizing deep reinforcement learning to learn from imbalanced data.The basic motivation is to design a reward function and the rules between the environment and agent to make the agent pay more attention on the minority class.The agent is able to learn the policy of imbalance classification by accumulating sufficient rewards.(2)A Deep Q Network(DQN)based imbalance classification model is proposed,which is named DQNImb,including the establishment of simulated environment,the definition of rules between the environment and the agent,and the designing of reward function.Specifically,the reward function produces relatively high rewards and penalties for the classification action of the minority,making the agent more concerned on the minority class.(3)We further evaluate the DQN based binary imbalance classification model through theoretical analysis and several experiments,including the superiority and robustness,the convergence of classification policy,the reasonability of reward function,and the reliability of model training.Experimental results demonstrate that our model outperforms existing known imbalance classification algorithms,and show strong generalization ability even if the data distribution is extremely imbalanced.(4)The DQN based imbalance classification model is also applied to imbalance multi-classification problems.Experimental results demonstrate that our model can also effectively solve the imbalance multi-classification problems.