Sparse PCA, variants, algorithms and applications

Laurent El Ghaoui
University of California, Berkeley (UC Berkeley)

In recent years there has been strong interest in modifying classical data dimensionality reduction algorithms, such as principal component analysis (PCA), in order for the result (say, a direction of high data variance) to be more interpretable, for example, sparse. In this talk we review some applications of this method to large data sets, highlighting the non-obvious fact that sparse versions can be easier to apply than their classical counterparts. We also explore the application of the row-by-row method (block coordinate descent) to these sparse dimensionality reduction problems.

Back to Workshop V: Applications of Optimization in Science and Engineering