Abstract:

There is a basic assumption in many existing data mining and machine learning techniques，that training and test data must be governed by the same distribution.However，this assumption does not hold in many cases，then traditional machine learning methods not aware of the difference of distribution may fail.This paper proposes a novel transfer-learning algorithm called DRTAT，which dynamically regroups the primary training data sets and eliminates the redundancy data timely，then makes classifiers ensemble.The experiments are performed on many text data sets and the UCI benchmark data sets，and DRTAT is compared with TrAdaboost algorithm，the results show the superiority of DRTAT.