Local regression transfer learning for users' personality prediction

Abstract

Some research has been done to predict users' personality based on their web behaviors. They usually use supervised learning methods to model on training dataset and predict on test dataset. However, when training dataset has different distributions from test dataset, which doesn't meet independently identical distribution condition, traditional supervised learning models may perform not well on test dataset. Thus, we introduce a new regression transfer learning framework to deal with this problem, and propose two local regression instance-transfer methods. We use clustering and k-nearest-neighbor to reweight importance of each training instance to adapt to test dataset distribution, and then train a weighted risk regression model for prediction. We perform experiments on the condition that users dataset are from different genders and from different districts, and the results indicate that our methods can reduce mean square error about 30% to the most compared with non-transfer methods and be better than other transfer method in the whole. © 2014 Springer International Publishing.