Study Of Class Imbalance Learning Based On Extreme Learning Machine

As one of the Top10 challenges in data mining field,the class imbalance problem has been focused by both academia and industry.As its name indicates,a data set is called class imbalance means the number of one class is far more than the others in this data set.This problem could generally result in the bias of the classification boundary towards the majority class for a traditional classification model,which further leads to the decrease of the classification performance.In recent two decades,a variety of solutions to deal with this problem have been proposed,but they almost ignore a common problem that the samples' prior distribution might affect the performance of the classifier.To address the problem above,we study the impact of class imbalance on the performance of classification model.Then,from the aspect of extracting precise information from the prior data distribution,two new models which respectively address binary-class/multi-class imbalance problem and one-class classification problem are presented.The main contributions of this thesis include the following two aspects:1)In a traditional binary-class or multi-class imbalance classification problem,the decision boundary always biases towards the majority class,further leading to the decrease of prediction accuracy of the minority class.To deal with this problem,this thesis presents an algorithm called CPWELM(Coupling Prior distribution Weighted Extreme Learning Machine),which extracts information from the prior distribution information of the dataset e.g.,imbalanced ratio,class overlapping size,noise and outliers,etc.Then,we integrate these information into a fuzzy weight matrix to optimize the hyperplane and promote the generalization performance.We conduct experiments on 46 imbalanced datasets to verify the feasibility and effectiveness of the proposed algorithm.The results indicate that the proposed algorithm generally performs better than the state-of-the-art ones.2)One-class classification is an extreme case in class imbalance problem,which means that the instances of the minority class is difficult to be obtained,causing the boundary only be determined by the target class.The applications of one-class classification includes anomaly detection,novelty detection,etc.The traditional one-class classifiers have drawbacks as poor adaptability for data distribution and inaccurate feature description.To address the problem mentioned above,this thesis presents an approach named fuzzy one-class extreme auto-encoder(FOCEAE).Combining with fuzzy extreme learning machine(FELM),K-nearest-neighbors(KNN)is used to estimate the relative density of each training object,then the fuzzy weight matrix is generated and integrated into FELM,further realizing the accurate description of the instances belonging to the target class.In particular,FELM is trained in the form of the auto-encoder and determine the classes which test samples belonging to by ranking the reconstruct errors of all training instances.The experimental results indicate that in comparison with some conventional methods,FOCEAE not only has better ability of data description,but also can decrease the time-consumption to a large extent.