摘要：In many traditional machine learning algorithms, a major assumption is that the training samples and the test samples have the same distribution. However, this assumption does not hold in many real applications. In recent years, transfer learning has attracted a significant amount of attention to solve this problem. In this paper, a novel transfer learning method based on clustering analysis and re-sampling is proposed, which can correct different types of domain differences and does not need to estimate the different distribution directly. The method explores the data structure by clustering analysis, and then uses the obtained structure information to generate a new training set for target learning under a re-sampling strategy. To explore more data structure information and be more robust to data sets with various shapes and densities, the method introduces the fuzzy neighborhood membership degree to improve the performance of clustering analysis. It also applies the Gaussian kernel function to measure the similarities between samples to improve the reliability of the new training samples. The proposed method can transfer more usefi.il knowledge from the source domain to the target domain. Experimental results on toy datasets demonstrate that the proposed method can effectively and stably enhance the learning performance. Finally, the proposed algorithm is applied to the communication specific emitter identification task and the result is also satisfying.
关键词：Transfer learning;clustering analysis;density- based clustering;fuzzy neighborhood;communication specific emitter identification
