Human Activity Recognation Based On Ensemble Learning

Human activity recognition is an important research in artificial intelligence,which uses supervised machine learning to identify users'motion.Human activity recognition based on smart phone is particularly significant when those system save and analyze data in smart terminals.We combine ensemble learning and dimensionality reduction to increase classification accuracy because data in smart phone is much redundant.Random projection is an efficient and global dimensionality reduction method which keeps Lipschitz property with high probability.We add random projection to Adaboost methods before every iteration to increase ensemble diversity.The new method Ad_RFE outperforms Adaboost when base classifiers increase,which we explain by evaluation index Q statistics and P_Precise.Multi view ensemble method is a good choice when variables are redundant and required fewer base classifiers,in which an important question is how to partition feature space.Vipin Kumar and Sonajharia Minz propose OFSP method based on forward search.We add feature selection technique based on condition mutual information and backward search to OFSP algorithm.The new method outperforms OFSP in many cases.We also test the new algorithm on other UCI data sets.