K-means Based Transfer Learning Algorithm

Focused on the issue that most transfer learning methods ignore the intra-domain distribution structures of the target domain, an algorithm based on K-means (K-means Transfer Learning, KTL) is proposed to enhance the transfer learning performance of the classification algorithms. In view of the poor clustering behavior of K-means on non-convex data sets, the multi-core version of KTL is proposed by clustering data into more small clusters to better fit the non-convex distribution of data (MKTL, Multi-core K-means Transfer Learning). The experimental results show that MKTL achieves the best average accuracy in 3 datasets. Compared with the original methods (kNN, TCA, GFK, JDA), the performance of MKTL is improved by 2.5 similar to 12.8 percentage in high computational efficiency.