Line H. Clemmensen
Sparse Supervised Analysis
This talk considers sparse versions of supervised analysis for regression and classification. Sparsity and dimension reduction is of particular interest when data have many more variables than observations (large p, small n problems). It not only identifies variables of interest, but also in most cases obtains generalizable solutions despite the scanty sampling. This is a result of combining the l2- and l1- norm of the parameter estimates. This approach in general works well when there exists a high degree of correlation amongst the covariates. We will consider sparse regression, sparse discriminant analysis, and sparse partial least squares analysis. The methods will be illustrated with applications in genome research and (spectral) image analysis.