Research On Imbalanced Classification And Multimodal Classification In Broad Learning System

With the rapid development of machine learning and computational intelligence,the research achievements of neural networks for massive high-dimensional data are constantly emerging.In 2018,Professor C.L.Philip Chen of South China University of Technology(SCUT)proposed a shallow network structure of Broad Learning System(BLS).The input BLS data is effectively represented by feature mapping and sparse dictionary to generate mapping nodes,and then the mapping nodes are mapped to generate enhancement nodes.The output is generated by the optimal connection matrix of ridge regression generalized inverse,which is composed of the mapping nodes and the enhancement nodes.Compared with traditional structures such as Convolution Neural Networks(CNNs)and Deep Believing Network(DBM),BLS not only has fewer super parameters and higher computational efficiency,but also realizes incremental learning of data and nodes.Therefore,more and more researchers pay attention to this novel network.In this thesis,two classification problems,imbalanced data and multi-modal data,are studied deeply in broad learning system.(1)For class imbalanced data,a new Weighted BLS(Adaboost-WBLS)based BLS and Adaboost is proposed.Broad learning system is a feed-forward single hidden layer algorithm network.When processing the data with unbalanced categories,the features of the few categories extracted are still insufficient compared with the majority of the categories,leading to the poor performance of data recognition and classification of the few categories.The mapping relationship between the output layers,the output layers and the single hidden layers of BLS is very simple,so adding weights to the corresponding features of the single hidden layer is a common way to re-balance categories of model.Based on KKT(KarushKuhn-Tucker)condition,this paper deduces the optimization process of adding diagonal matrix formal weight to the width learning system,and verifies the inhibitory effect of adding diagonal weight on BLS model errors.Weighted BLS(W-BLS)has a strong similarity with Adaboost integration process in order to realize online update with added weights.Therefore,Adaboost is used to integrate W-BLS,and the imbalanced data classification algorithm of Adaboost-WBLS is proposed.The dynamic update of weights is realized through iteration to obtain weights that are more consistent with data characteristics,so as to further improve the recognition ability of the integrated model for minority classes.When the weights are initialized,the initialization strategy based on category information is adopted to make the model have higher integration training efficiency.In the process of weight updating,different regularization updating methods are adopted for different classes to retain the characteristics within the data class and increase the distinction between classes.The ablation experiment and comparison experiment are conducted on the Adaboost models and BLS models respectively.Experimental results prove that the model has a significant improvement in the recognition ability of minority classes.(2)Cost-sensitive Stacked BLS(Cs-Stacked BLS)are proposed to solve the problem that the output of the stack broad learning system is not clear and the stack process lacks cost control.Stack broad learning system is a model that approximates stack residuals to achieve the highest test accuracy.However,when dealing with the classification task,there are still some problems,such as insufficient classification decision differentiation and insufficient stack process cost analysis.In order to enhance the decision differentiation,this paper proposes CS-stacked BLS based on sequential three-way decisions,so as to realize the comprehensive evaluation of decision cost,accuracy and time cost in the stack process.Experiments on four commonly used classification datasets show that CS-stacked BLS is an effective network that comprehensively considers decision risk and process uncertainty.(3)For multi-modal datasets,a Cost-sensitivy Voting Stacked BLS(Cs-V Stacked BLS)multi-modal classification algorithm for decision fusion is proposed.The complex classification process of multi-modal data is a process of effective extraction and then fusion of single modal data features.As an effective algorithm for multi-index evaluation,Cs-Stacked BLS has the advantages of good generalization performance,stable decision-making process and result.However,the process of feature extraction is redundant,so the decision fusion method based on sensitive weighted voting is adopted to make it capable of multi-modal data classification.Compared with same algorithms the performance of the new algorithm for multi-modal classification of complex problems is verified.