A0781
Title: Regularized reduced-rank regression for structured output prediction with vector-valued functions
Authors: Kun Cheng - Beijing Jiaotong University (China) [presenting]
Abstract: In predicting multiple response variables from the predictor variable, the reduced-rank regression (RRR) is an effectively linear method that implies that the matrix of regression coefficients is of lower rank. The focus is on exploiting RRR with a reproducing kernel Hilbert space (RKHS) approach. It models the multiple response variables as a linear combination of a few vectors with coefficients being (possibly nonlinear) functions of the predictor variable. In the underlying setting, coefficient functions are chosen from RKHS, while in RRR, they are restricted to linear functions. A set of solutions in RKHSs is characterized by the help of cross-covariance operators in RKHSs. Moreover, regularized estimators are constructed, and estimation errors are bounded by mild assumptions. A convergence rate for estimation errors is established. Simulations and real data analysis are provided to illustrate the efficiency of the proposed method.
