A0771
Title: Ultrahigh-dimensional discriminant analysis and its application to gene expression data
Authors: Li-Pang Chen - National Chengchi University (Taiwan)
Jou Chin Wu - National Chengchi University (Taiwan) [presenting]
Abstract: Discriminant analysis has been a commonly used strategy to handle classification with binary or multi-label responses. Under multivariate normal distributions of covariates, a linear or quadratic discriminant function can be derived, which is used as a boundary to classify subjects. In the discriminant function, the estimation of the precision matrix, which is defined as the inverse of the covariance matrix, is a crucial issue. While a large body of estimation methods is available to estimate the precision matrix, most methods not only fail to handle ultrahigh-dimensionality in the sense that the dimension of variables is extremely larger than the sample size but also require longer computational time. In addition, in the presence of nonlinear dependence among variables, existing methods may falsely miss the detection of dependence. To tackle those challenges, the model-free feature screening method is extended to reduce the dimension of variables and detect possibly nonlinear pairwise dependence structures among variables. After that, the graphical lasso and joint graphical lasso methods are adopted to estimate the precision matrix and then implement the estimator to the discriminant function. Numerical studies are conducted to assess the prediction performance of the proposed method.
