Imbalance Classification Methods Based On Transfer Learning And Their Application

Data collection and annotation require a lot of resources in practical applications.And many collected datasets are imbalanced.The classification of imbalanced data has great significance in many fields such as fault detection,medical diagnosis,and credit card fraud detection.Traditional machine learning classification methods require that the training data and testing data must satisfy independent and identical distribution.And many methods are based on the assumption that the dataset is balanced.When the training data and the testing data do not meet the above assumptions and the amount of labeled data is very small,the traditional classification learning method is difficult to build a reliable model because of the scarcity of labeled data and the tendency of majority classification.In this case,transfer learning and imbalanced classification are very necessary.This thesis combines the imbalanced characteristics of data to complete the work from the following three aspects: model,instance and feature.(1)A parameter-based transfer learning algorithm for imbalanced classification.For the problem that training data and testing data are imbalanced data and their corresponding parameter spaces are different,an imbalanced classification algorithm based on parameter transfer is given.From the perspective of model construction,this algorithm solves the problem of poor classification ability of traditional machine learning algorithms in imbalanced dataset.Moreover,the algorithm aligns the parameter space of training data and testing data.As a result,the structure of the classifier is improved.The core of this algorithm is to build a transformation matrix,so that the training data and testing data share the same parameter space.This algorithm not only has good classification performance on the public datasets,but also performs well in the detection of micro-seismic.(2)An instance-based transfer learning algorithm for imbalanced classification.Traditional transfer classifiers cannot obtain high classification evaluation performance when the dataset is imbalanced,so we construct a weighted scheme that it sensitive to the misclassification of minority samples.This algorithm aims to improve the influence of minority samples in classification.And it can reduce the misclassification cost of the minority class samples while ensuring the accuracy of the majority class samples classification.At the same time,the algorithm also uses boosting methods.It dynamically adjusts the weight of the weak classifier through classification evaluation metrics.In the end,we obtain an imbalanced data classifier with strong robustness and high classification performance.Experiments on a variety of public datasets and micro-seismic datasets show that the effectiveness of the proposed method.(3)A feature-based transfer learning algorithm for imbalanced classification.In order to solve the problem of the large difference in feature distribution between training data and testing data,a feature-based transfer learning algorithm for imbalanced classification is proposed.The algorithm gives the two definition that are named as Credibility of Feature and Contribution of Feature to Transfer to dynamically evaluate the process of feature transfer learning.The algorithm also reduces the complexity of the neural network classifier by reducing redundant features.As a result,the difference between the feature distribution of the source domain and the target domain is reduced.Finally,the performance metric of the classifier has been effectively improved.The performance of this algorithm is verified by experiments on both public datasets and micro-seismic datasets.The thesis includes 20 figures,16 tables and 114 references.