论文部分内容阅读
针对多标记迁移学习中源领域与目标领域的特征分布差异会导致源领域数据无法被目标领域利用的问题,提出了一种基于最大均值差异的多标记迁移学习算法(Multi-Label Transfer Learning via Maximum mean discrepancy,M-MLTL),算法通过分解关系矩阵构造共享子空间,并采用最大均值差异(maximum mean discrepancy)作为评价指标,最小化子空间特征的分布差异,从而使源领域与目标领域的特征分布尽可能相似.多标记图像分类实验的结果表明,新算法比同类算法有更高的精度和计算效率.
To solve the problem that the difference of feature distribution between source and target domains in multi-label migration learning can lead to the inability of source domain data to be utilized by the target domain, a Multi-Label Transfer Learning via Maximum Maximum Multi-label Transfer Learning mean discrepancy and M-MLTL). The algorithm constructs the shared subspace by decomposing the relation matrix and adopts the maximum mean discrepancy as the evaluation index to minimize the distribution differences of the subspace features, so that the characteristics of the source domain and the target domain The distribution is as similar as possible.The results of multi-label image classification experiments show that the new algorithm has higher accuracy and computational efficiency than the similar algorithms.