MMDE

regarded as a MMD-based feature extraction method

has been successfully used.However

when the feature extraction problems of the original input space have been solved

the MMDE lacks the suitability to some extent.Therefore

we propose Maximum Local Weighted Mean Discrepancy(MLMD)by integrating the theory and technique of local learning methods.The measurement considers fully the internal local structure between domains;at the same time

the global distribution discrepancy can be reflected by the local distribution discrepancy.We also

based on the above measurement

propose Maximum Local Weighted Mean Discrepancy Embedding(MWME)

which not only fulfills transfer learning task but also has certain local learning capability.The MWME can complete traditional feature extraction as well as domain adaptation learning in two domains whose distributions are different but relative

thus indicating its better robustness and adaptation.Tests show the above-proposed advantages of the MLMD criterion and the MWME method.