Distributed tuning of machine learning algorithms using MapReduce Clusters
Yasser Ganjisaffar, Thomas Debeauvais, Sara Javanmardi, Rich Caruana, Cristina Videira Lopes · 2011
Obtaining the best accuracy in machine learning usually re-quires carefully tuning learning algorithm parameters for each problem. Parameter optimization is computationally challenging for learning methods with many hyperparame-ters. In this paper we show that MapReduce Clusters are particularly well suited for parallel parameter optimization. We use MapReduce to optimize regularization parameters for boosted trees and random forests on several text prob-lems: three retrieval ranking problems and a Wikipedia van-dalism problem. We show how model accuracy improves as a function of the percent of parameter space explored, that accuracy can be hurt by exploring parameter space too aggressively, and that there can be significant interaction between parameters that appear to be independent. Our results suggest that MapReduce is a two-edged sword: it makes parameter optimization feasible on a massive scale that would have been unimaginable just a few years ago, but also creates a new opportunity for overfitting that can reduce accuracy and lead to inferior learning parameters.