模糊近邻密度聚类与重采样的迁移学习算法

Transfer Learning with Fuzzy Neighborhood Density Clustering and Re-sampling

传统机器学习要求训练样本和测试样本具有相同分布的假设在实际应用中难以满足,为解决这种问题,迁移学习的研究近年来逐渐兴起。其中,基于聚类分析与重采样的迁移学习框架不需要直接估计域分布,且能够修正不同类型的域间差异,但其所采用的聚类算法对参数选择的鲁棒性及不同分布数据的适应性较差,并不能很好地适用于挖掘数据结构信息。为此,该文提出一种基于模糊近邻密度聚类与重采样的迁移学习算法。该方法对不同分布形状和密度的数据具有较好的鲁棒性并能够发现更多的近邻结构信息,能够从源域中迁移更多的有用知识用于目标域的学习。在公共数据集上的实验结果表明所提出的迁移学习方法具有更好的性能。

In many machine learning algorithms, a major assumption was that the training samples and the test samples had the same distribution. However, this assumption did not hold in many real applications. In recent years, transfer learning had attracted a significant amount of attention to solve this problem. Among these methods, an effective algorithm based on clustering analysis and re-sampling could correct different types of domain differences and did not need to estimate the dif- ferent distribution directly. As the critical part, the original clustering algorithm was not good enough at data structure ex- ploration due to its poor robustness on the data with various shapes and densities. In this paper, a new transfer learning al- gorithm based on fuzzy neighborhood density-based clustering and re-sampling was proposed, which was more robust to datasets with various shapes and densities, and could explore more data structure information. With the better explored data structure information, the proposed method could transfer more useful knowledge from source domain to target domain. Val- idation of the proposed method was performed with extensive experiments. Results demonstrate that the proposed method can more effectively and stably enhance the learning performance.