Agoston Mihalik - CMIC/WEISS joint seminar series

Title - Regularized canonical correlation analysis: a unifying concept for CCA/PLS/RRR and practical considerations

 Abstract

Recently, there has been a growing interest in using multivariate approaches such as Canonical Correlation Analysis (CCA), Partial Least Squares (PLS) and Reduced Rank Regression (RRR) to investigate associations between multiple views/sources of data e.g. brain imaging and genetics or brain imaging and behaviour. To control the complexity of the CCA model and the risk of overfitting, a regularised version of CCA (RCCA) was proposed by Hardoon et al. 2004. In this talk, I will demonstrate that PLS, CCA and RRR can be seen as specific cases of RCCA. Moreover, I will elucidate how these models are linked to multivariate multiple linear regression or general linear model (GLM). I will present two popular extensions of RCCA: kernel CCA (KCCA) and sparse PLS (SPLS). KCCA improves the computational efficiency of the model, whilst SPLS facilitates the interpretability of the results. I will discuss how to improve the generalizability and stability of these regularized CCA/PLS models using a machine learning framework. Finally, I will present a recently developed toolkit which includes the discussed models, provides data-driven selection of the regularization parameters, evaluates the statistical significance of the results and can be easily run using MATLAB.