Machine Learning In Complex Hypothetical Scenarios

With the rapid development of computer technology,techniques for data obtaining and storing are improved fast.These data often originate from complex sources,and also contain noise information and redundant information.Using these data to establish a reliable model has been a hot topic in machine learning.To deal with label noise,r NDA algorithm is presented by Jakramate Bootkrajang via introducing a flipping probability into Normal Discriminant Analysis(NDA).For complex originated data,input features could not be represented by a simple distribution.In order to solve this problem,Gaussian mixture model is brought into r NDA algorithm and a robust Discriminant Analysis based on Gaussian mixture model(GMDA)is proposed.Experiments show the newly proposed algorithm behaves well in classification,and achieves better performance compared with original NDA and r NDA.The number of mixture components in GMDA is given in advance,and an improper specified number will affect model's predict performance.A Dirichlet Process is considered to adaptively solve this problem.The Dropout method applied to Deep Learning is brought into classical classification algorithm to deal with the noise and redundant information in input features.Take advantage of Dropout method to select feature subsets,an ensemble classifier based on Dropout feature selection is proposed.Experiments show the proposed ensemble classifier has a significant improvement on classification performance,and obtains outstanding predictions on datasets containing label noise.