Study Of Classification Problems Based On Sparse Representation And Ensemble Learning

Classification problem is a general problem in the real world, as well as the one ofcore problems in the machine learning community. Driven by problems in the real world,classification problem has been extended from single-instance single-label classification(traditional supervised classification) to multi-label classification, multi-instanceclassification, and multi-instance multi-label classification. The above variousclassification problems are new challenges for researchers from machine learningcommunity.Sparse representation and ensemble learning have sound theoretical foundations,and are strong tools for solving classification problem. They have manifested goodperformance in many applications. In order to solve the above various classificationproblems, this dissertation focus on the concrete classification problems ofsingle-instance single-label hyperspectral remote sensing image, multi-label image,multi-label gene, multi-label Web page, multi-instance image, and multi-instancemulti-label image. In order to improve the performance of classification, some newmethods are proposed based on sparse representation and ensemble learning. The mainresearch fruits achieved in this dissertation can be summarized as follows:1. A novel classification method of hyperspectral remote sensing image is proposedbased on sparse representation feature and spectral information feature. First, adictionary is obtained based on hyperspectral remote sensing image data and machinelearning method, and then the sparse feature of each pixel is calculated according to thedictionary. Finally, random forest is used to respectively classifying sparserepresentation feature and spectral information feature, and the ensemble ofclassification results is used for prediction. The experimental results conducted onhyperspectral remote sensing image data indicate that the proposed method has betterclassification results than those of based on spectral information feature and based onsparse representation feature.2. A novel multi-label classification method based on sparse representation isproposed. Firstly, the training samples are used as a dictionary, and the test sample istreated as a linear combination of training samples in the dictionary, the sparserepresentation coefficients are obtained based on l1-minimization method. Thendiscriminating information of sparse representation coefficients is utilized to calculatemembership function of the test sample belonging to labels. Finally, labels are rankedaccording to membership function and the test sample is assigned to labels by using rank result. Extensive experiments conducted on multi-label data show that theproposed method achieves better results than other works in the literatures.3. A multi-label classifier ensemble method based on random subspace is proposed.The multi-label base classifier in the multi-label ensemble system is constructed byrandomly selecting subsets of components of the whole feature, and the result ofclassifier ensemble is used for prediction. The experimental results on the multi-labeldata demonstrate that the performance of the proposed method is better than those ofsingle multi-label classifier.4. A multi-instance image classification method is proposed based on sparserepresentation and ensemble learning. First, a dictionary is learned based on all theinstances in the training bags, and the sparse representation coefficients of each instanceare calculated according to the dictionary; second, a bag feature vector is computedbased on all the sparse representation coefficients of instances in the bag. Thus,multi-instance classification problem is transformed into traditional supervisedclassification problem that can be solved by well-know traditional supervisedclassification methods. In order to improve classification performance, the componentclassifiers are obtained by repeatedly using the above method with dictionaries ofdifferent sizes, and the result of classifier ensemble is used for prediction. Experimentalresults on multi-instance image data demonstrate the superiority of the proposed methodin classification accuracy as compared with state-of-the-art multi-instance classificationmethods.5. According to the idea of degeneration, a novel multi-instance multi-label imageclassification method is proposed based on sparse representation and classifier ensemble.First, the bag feature of multi-instance multi-label image is computed by sparserepresentation method based on dictionary learning, and multi-instance multi-labelimage classification problem is transformed into multi-label classification problem.Second, multi-label classification problem is further transformed into traditionalsupervised classification problem, therefore traditional supervised classificationmethods can be used to solve it. In order to improve classification performance, manydiversity individual classifiers are constructed by repeatedly using the above methodwith dictionaries of different sizes, and the result of classifier ensemble is used forprediction. Experimental results conducted on multi-instance multi-label image datashow that the proposed method is superior to the state-of-art methods in terms ofmetrics.