Abstract
Linear classification is a useful tool for dealing with large-scale data in applications such as doc-ument classification and natural language process-ing. Recent developments of linear classification have shown that the training process can be effi-ciently conducted. However, when the data size ex-ceeds the memory capacity, most training methods suffer from very slow convergencedue to the severe disk swapping. Although some methods have at-tempted to handle such a situation, they are usually too complicated to support some important func-tions such as parameter selection. In this paper, we introduce a block minimization framework for data larger than memory. Under the framework, a solver splits data into blocks and stores them into separate files. Then, at each time, the solver trains a data block loaded from disk. Although the framework is simple, the experimental results show that it ef-fectively handles a data set 20 times larger than the memory capacity.