Deep Transfer Learning For Unbalanced Data Classification

Unbalanced classification has always been a hotspot in artificial intelligence.Most of traditional classification algorithms take the basic balance of data distribution as the premise,and the classification accuracy of the sample as a goal,which does not apply to unbalanced classification.Deep network can effectively learn the features of data.Transfer learning can use the data of the existing relevant tasks and domains processing target tasks and the problems of target domains.The combination of deep learning and transfer learning is known as the Deep Transfer Learning,DTL.Deep Transfer Learning uses the auxiliary data in the relevant domain to establish the deep network model and study the features of the relevant data.By transferring the auxiliary deep network structure and features which are conducive to the unbalanced classification,the classification ability of the target network to unbalanced data will be improved,the time of training model will be saved,and the generalization of the model will be promoted.The specific work done in my thesis is as follows:First,my thesis proposes a binary transfer learning algorithm,UTr A,for unbalanced data.Based on the TrAdaboost algorithm,UTrA calculates the weights of weak classifiers according to the area under Precision-Recall curves of two classes,and applies different weight update strategies on the instances of different classes.Secondly,my thesis proposes a binary classification,the Ensemble Unbalanced Deep Transfer algorithm,EUDT,for unbalanced data.The EUDT algorithm trains auxiliary data using deep network and transfers the network structure and the features to the target data.The target network is retrained by the improved average precision error function APE and the average precision cross entropy loss function APCE,which can effectively learn the features of the target data and improve the ability to identify minority class.The ensemble of transfer classifiers can save time cost on choosing transfer mode,alleviate the excessive tendency to majority of transfer classifiers,and get meaningful result of unbalanced classification.When evaluated the unbalanced classification,since AUC is insensitive to changes in class distribution,it will be more accurate to take AUC value combined with G-mean and BER.The results of these three metrics indicate that,the UTrA and EUDT algorithms achieve better performance for imbalanced data,it focuses more on minority instances while keeps the accuracy of majority classification,even in complex situation.