72 Partial Least Squares
72.1 Partial Least Squares
We can think of partial least squares (PLS) as a variant of PCA. What makes it different than PCA is that we are using the outcome to guide the method, unlike what happens in classical PCA.
One maximizes the variance of each principal component, regardless of whether the principal component has any relation to the outcome. This means that we could throw away a highly predictive low-variance component by using traditional PCA.
PLS works by extracting components that maximize the covariance between the predictors and the outcome. This makes it so it has the same structure as PCA, but the loadings will be different to account for this additional constraint.
72.2 Pros and Cons
72.2.1 Pros
- Can give better results than PCA
72.2.2 Cons
- Computationally slower