Sparse PCA with Multiple Components

Sparse PCA with Multiple Components

As the dimension of data sets increases, analysts often rely on principal component analysis (PCA) to summarize data via a small number of informative principal components (PCs). Sparse PCA makes those directions easier to interpret by using only selected variables, but computing several sparse components at once creates a problem: standard one-at-a-time methods lose the orthogonality that makes PCA useful in practice. In “Sparse PCA With Multiple Components,” Ryan Cory-Wright and Jean Pauphilet develop optimization-based methods that choose multiple sparse principal components simultaneously. Their approaches combine semidefinite relaxations, Lagrangian decompositions, and a new combinatorial upper bound to produce sparse, orthogonal components along with certificates of near optimality. Across real and synthetic data sets, the methods deliver high-quality solutions at practical scales, with average bound gaps around 3% for real-world instances with hundreds or thousands of features. The work gives practitioners a principled way to obtain interpretable low-dimensional representations without sacrificing the structural guarantees of PCA.