B1120
Title: Linear discriminant analysis with high dimensional mixed variables
Authors: Zhongqing Yang - The Hong Kong Polytechnic University (Hong Kong) [presenting]
Binyan Jiang - The Hong Kong Polytechnic University (Hong Kong)
Cheng Wang - Shanghai Jiao Tong University (China)
Chenlei Leng - University of Warwick (United Kingdom)
Abstract: With the rapid development of modern measurement technologies, datasets containing both discrete and continuous variables are more and more commonly seen in different areas. The dimensions of the discrete and continuous variables are often be very high. Though discriminant analysis for mixed variables under the traditional fixed dimension setting has been well studied since the 80's, promising approaches taking into account both the high dimensionality and the mixing nature of the data sets are still missing. We aim to develop a simple yet useful classification rule that addresses both the high dimensionality and the mixing nature of the variables simultaneously. Our framework is built on a location model, under which we further propose a semiparametric formulation for the optimal Bayes rule. We show that the optimal classification direction and the intercept in the optimal rule can be estimated separately. Efficient direct estimation schemes are then developed to obtain consistent estimators of the discriminant components. Asymptotic results on the estimation accuracy and the misclassification rates are established, and the competitive performance of the proposed classifier is illustrated by simulation and real data studies.
