Distributed adaptive lasso penalized generalized linear models for big data.

Adaptive lasso penalized generalized linear models (GLMs) are a powerful tool for analyzing the high-dimensional sparse data where the classical linear or normal assumption is not met. In non-distributed environments, the estimation problem of adaptive lasso penalized GLMs is often solved by the coordinate descent based algorithm developed in Friedman, Hastie, and Tibshirani (2010), which has been well implemented in the R package glmnet. However, when applied to distributed big data, this algorithm is usually inflexible or even infeasible due to its non-parallel implementation, especially when the communication costs between the central and local machines are expensive, or the storage and computing capabilities of the central machine are insufficient. In this paper, we propose a new method, QAGLM-alasso, for the adaptive lasso penalized GLMs problem in distributed big data by applying the quadratic approximation representation of GLMs, and further develop a path-following algorithm for its estimation based on the Least Angle Regression (LARS). Theoretical analyses show that, under mild regularity conditions, the QAGLM-alasso enjoys the oracle property, and the obtained estimator is asymptotically equivalent to the original adaptive lasso. Simulation studies demonstrate that the new algorithm has similar estimation accuracy with glmnet, but is significantly faster than glmnet in distributed environments. We further illustrate the practical performance of the proposed method by analyzing a supersymmetric (SUSY) benchmark data set. [ABSTRACT FROM AUTHOR]