Summary: | 碩士 === 國立東華大學 === 應用數學系 === 100 === The analysis of high dimension and low sample size is emerging as a problem of theoretical and practical importance for statistics and machine learning in recent years.The performance of Fisher Discriminant Analysis (FDA) deteriorates considerably for data sets of high dimension and low sample size. Under an asymptotic framework, however, Bickel and Levina (2004) proves that the Linear Discriminant Analysis (LDA) coupled with naive Bayes assumption outperforms the ordinary LDA. Through analysis of simulations and benchmark data sets, we try to understand the performance of LDA with naive Bayes assumption in more practical settings. Besides naive Bayes assumption, we also discuss other different regularization in FDA, and look at their performance.
|