Research On Reinforcement Learning Transfer Algorithm For Imbalanced Data Classification

With the rapid progress of technology,the output of data has reached an unprecedented level and these data can be collected and stored in large quantities.Therefore,the demand for data processing in different industries is also increasing.In the fields of machine learning and data mining,classification techniques are widely used,which can help us quickly and accurately identify complex information.These techniques can be implemented through various machine learning methods.Data classification technology has become an important tool for intelligent processing,which can effectively help us better identify and predict various categories of data sets,thereby greatly improving the efficiency and accuracy of data analysis.Traditional classification algorithms can provide accurate and effective classification prediction results when dealing with balanced data category distribution.However,real-life data usually exhibits imbalanced distribution,and traditional classifiers perform poorly in classifying imbalanced data sets.Scholars have conducted existing research on imbalanced data classification mainly from three aspects: data preprocessing,feature selection,and algorithm optimization.Regarding the classification problem of imbalanced data,the dataset not only suffers from imbalanced sample distribution but also insufficient volume to support the training of a good model.Furthermore,training a model from scratch tailored to the current dataset incurs substantial time and financial costs.Adopting appropriate auxiliary datasets to transfer information to the imbalanced target dataset can effectively address the issue of poor model performance caused by the imbalanced distribution and scarce sample quantity of the target dataset.These datasets can assist us in transferring knowledge from other collections to the imbalanced collection,thereby improving the accuracy and reliability of the model.This paper proposes a reinforcement learning transfer algorithm tailored to imbalanced data classification.First,a deep convolutional generative adversarial network is employed as an oversampling method,generating effective minority-class samples to reduce the gap between majority and minority classes.Then,a deep Q-network is used as the classifier’s imbalanced data classification model,in which the reinforcement learning reward function is added to the classification model,and optimization is utilized to increase the agent’s cumulative reward and enhance classification accuracy.Finally,the deep Q-network is utilized as the imbalanced data classification model for transfer learning,ensuring that the model performs well on another similar dataset without negative transfer.The transfer is implemented separately for structured and unstructured datasets in both the source and target domains,making the deep Q-network imbalanced data classification model suitable for various tasks.