Classification Methods Based On Broad Bayesian Neural Networks

Classification is an important research direction in the field of machine learning.As an optimization approximation method,the Variational Bayesian Inference algorithm has achieved much applications in the fields of statistics and control science in recent years because of its good model scalability and data adaptability.In this paper,considering the Broad Learning theory,we propose a variant of Bayesian Neural Network architecture which feature mapping layer gradually broadened and prior distribution adaptively updated.The model is applied to the classification problems.The main contents of this thesis are listed as follows:Firstly,based on the Random Vector Functional Link Network,we design a Bayesian neural network classification model with its feature mapping layer gradually broadened.The feature expression is increased by gradually broadened network structure,and a local constant prior is introduced under the Gaussian prior distribution assumption to improve the training efficiency when the network structure is getting broadened.Experiments show that the proposed broad Bayesian neural network classifier can obtain good classification results and effectively resist the model robustness to harness the disturbed instances.In addition,the broaden structure and the prior updating algorithm can achieve a “stepwise” increase in classification accuracy and improve training efficiency.Secondly,this thesis extends its network structure based on the broad Bayesian Neural Network model.For the classification task of high-dimensional data,the rapid increase of parameter size leads to slower convergence of Bayesian neural network and the need for generalization of the model.Therefore,the model structure is extended: the sparse mapping method is adopted,and the enhancement layer is added to the network.Numerical simulations show that the extended broad Bayesian network model can not only obtain better classification results but also improve the learning speed of the network in the classification and recognition of image datasets.Finally,this thesis uses the broad Bayesian Neural Network model for the classification of imbalanced data.Aiming at rebalancing distribution of data in the quality of life diagnosis of dialysis patients,the method of generating disequilibrium data samples was designed,and the classification model of hemodialysis patients' quality of life based on broad Bayesian Neural Network was established.The experiments results show that the model can be effective.The quality of life of hemodialysis patients is divided into high quality group and inferior quality group,which provides a reference for doctors to formulate a treatment.