講座題目 | Envelope-based partial least squares | ||
主辦單位 | 數理與統計意昂2 | 協辦單位 | 應用統計系 |
講座時間 | 6月22日14:00-15:00 | 主講人 | Zhihua Su副教授 |
講座地點 | 行政樓1308室 | ||
主講人簡介 | Zhihua Su is Associate Professor in Department of Statistics at the University of Florida. She got her PhD in 2012 from University of Minnesota, Twin Cities. Her research interest includes Feature selection, Dimension reduction, Multivariate analysis, Bayesian statistics, Algorithm development, Software development, Applied statistics. | ||
講座內容簡介 | Partial least squares (PLS) is widely used in applied sciences an alternative method to ordinary least squares (OLS) for estimating the regression coefficients. It is known that PLS often has a better prediction performance compared to OLS, and the PLS algorithms can be adapted directly to the n < p case. Despite its popularity, the theoretical properties of the PLS estimator are largely unknown. As a result, it is hard to determine when PLS is better than OLS, what are the limitations for PLS and how to improve PLS. Cook et al. (2013) built a connection between PLS with a dimension reduction method called the envelope model. They showed that at the population level, PLS and the envelope model have the same target parameter, but they use different algorithms for estimation. This connection allows PLS to be studied in a traditional likelihood framework and facilitates model developments. We will address three issues of PLS in this context: variable selection, categorical predictors and scale invariance. | ||