|
Abstract:
|
Linear discriminant analysis (LDA) is one of the most commonly used methods for building a classification rule. Sparse LDA for high-dimensional classification often has weak performance due to small set of training samples. On the other hand, we often have a set of classification problems that are related to the target classification problem. This paper introduces transfer learning methods for LDA (TransLDA) that effectively utilize information from auxiliary data in order to build a better classification rule for the target study. The methods allow for both homogeneous and heterogeneous covariance matrices across different studies. In addition, an adaptive method is introduced that can identify the informative data sets in transfer learning. We show that under some assumptions, TransLDA has smaller error rates in estimating the discriminant directions and classification errors. We evaluate our proposed methods and compare its performance with standard LDA using simulations and observed smaller classification errors. We illustrate the proposed methods by building classification rule of colon cancer using gut microbiome data using data from different studies.
|
