knowledg transfer multipl model local structur map jing gao wei fan jing jiang jiawei han effect knowledg transfer classif algorithm depend differ distribut gener train exampl test exampl drawn task difficult train exampl domain test domain paper propos local weight ensembl framework combin multipl model transfer learn weight dynam assign model predict power test integr advantag learn algorithm label inform multipl train domain unifi classif model appli domain importantli previous propos method base learn method requir specif design transfer learn show optim local weight ensembl framework gener approach combin multipl model domain transfer propos implement local weight assign map structur model structur test domain weight model local consist neighborhood structur test experiment result text classif spam filter intrus detect data set demonstr signific improv classif accuraci gain framework transfer learn task newsgroup messag categor propos local weight ensembl framework achiev accuraci singl model predict correctli test exampl summari improv accuraci problem 