A Reliable Effective Terascale Linear Learning System

摘要

We present a system and a set of techniques for learning linearpredictors with convex losses on terascale data sets, withtrillions of features, (The number of features here refers tothe number of non-zero entries in the data matrix.) billions oftraining examples and millions of parameters in an hour using acluster of 1000 machines. Individually none of the componenttechniques are new, but the careful synthesis required to obtainan efficient implementation is. The result is, up to ourknowledge, the most scalable and efficient linear learningsystem reported in the literature. (All the empirical evaluationreported in this work was carried out between May-Oct 2011.) Wedescribe and thoroughly evaluate the components of the system,showing the importance of the various design choices. color="gray">