Early-Stopping Regularized Least-Squares Classification

We study optimization of the regularized least-squares classification algorithm, and proposes an early-stopping training procedure. Different from previous empirical training methods which separate model selection and parameter learning into two stages, the proposed method performs the two processes simultaneously and thus reduces the training time significantly. We carried out a series of evaluations on text categorization tasks. The experimental results verified the effectiveness of our training method, with comparable classification accuracy and significantly improved running speed over conventional training methods.