A1571
Title: Product-CCA: An ordering-robust canonical correlation analysis procedure without efficiency loss
Authors: Kenichi Hayashi - Keio University (Japan) [presenting]
Hung Hung - National Taiwan University (Taiwan)
Abstract: Canonical correlation analysis (CCA) is a fundamental statistical tool for extracting shared information from two sets of variables, but its performance can be highly sensitive to outliers. Although robust methods for CCA have been studied, the literature remains less developed than that on principal component analysis (PCA), a closely related and simpler method. Recently, a random-partition-based approach called product-PCA was proposed that improves outlier-robustness while preserving the estimation efficiency of conventional PCA in the absence of outliers. Motivated by this idea, Product-CCA (PCCA) is proposed, a new CCA procedure based on random partitioning of the data. PCCA improves the outlier-robustness of conventional CCA without directly downweighting observations suspected to be outliers. This property enables PCCA to share the same asymptotic efficiency as CCA, an appealing property in real applications. The construction of PCCA, its theoretical considerations, and results of numerical experiments are presented.
