Speaker: Song Cai (Carleton University)

Title: Memory-efficient computing algorithms for L1-type regression estimators based on big data

Date: November 28, 2014

Time: 1:30 pm – 2:30 pm

Room: 4351 HP (Macphail room)

ABSTRACT: The L1-based regression estimators, such as those in least absolute deviation regression, quantile regression, and rank regression, are usually preferred over the least square estimator because they are more robust to outliers in the data while having decent efficiency. However, when sample size is large, the standard technique for computing these estimators, the linear programming, requires much more computer memory than that usually found in modern personal computers at, say, 16 gigabytes. This problem becomes even more severe for big data, one of the most challenging topics in contemporary statistical research. In this work I aimed at making the computation of the L1-based estimators feasible, and here I present two subgradient-based optimization algorithms that uses small amount of computer memory while still converges reasonably fast.