Bicoordinate Descent for the LASSO

July 2015
with In Song Kim and Marc Ratkovic; Working Paper

Search Terms: 

We propose an estimator for the LASSO that iteratively optimizes the coefficients in pairs. In addition to improving efficiency by coordinating the updates of the the paired variables, our algorithm affords insights into the nature of the LASSO problem. Our method outperforms the popular glmnet algorithm in all but high-K low-N settings, executing increasingly better as N increases.