科研之友
微信
新浪微博
Facebook
分享链接
摘要
Variable selection in high dimensional data is a challenging problem due to the exponential number of variable combinations, and Markov Chain Monte Carlo (MCMC) methods represent the state of the art to solve it. With genomics data this problem becomes even more difficult because there are generally more dimensions (variables) than points (records) leading to slow convergence and numerically unstable solutions. On the other hand, despite many alternative prototypes and languages, R remains a popular system to compute machine learning models. Unfortunately, R can be particularly slow with heavy matrix computations and the high number of iterations required by MCMC methods. Moreover, making R scale to large matrices, possibly beyond RAM, requires careful system integration. Recently, array DBMSs have opened the possibility of manipulating matrices of unlimited size. With such motivation in mind, we present algorithmic optimizations to accelerate the computation of variable selection in linear regression with the Gibbs sampler, a fundamental MCMC method. Such optimizations have the potential to accelerate other models. We study how to leverage the speed and scalability of the array DBMS to exploit our optimizations in R. We present a comprehensive experimental evaluation to assess time efficiency and model quality with a cancer data set containing RNA and miRNA variables to predict survival time. We show our optimized algorithm combining DBMS and R processing is significantly faster than R alone. We show our system allows fast joint analysis of RNA and miRNA variables, instead of analyzing them separately. Finally, we confirm our algorithm finds medically significant variables already identified in the biomedical literature. Our optimized MCMC method for the array DBMS can be easily called from R, leaving the final model within R runtime in RAM for further interpretation.
- 出版日期2016-3
