局部标记关系的多标记迁移学习算法

Multi-label Transfer Learning by Exploiting Local Label Correlations

现有多标记迁移学习主要利用全局标记关系信息,由于源领域与目标领域的标记关系存在差异,源领域中全局标记关系不适合于目标领域.本文提出一种局部标记关系的多标记迁移学习算法,该算法通过对样本的标记进行聚类和最小化联合损失函数,可以有效的挖掘领域间共享的局部标记关系,对应的局部关系编码可以作为样本的辅助特征从而提高模型性能.图像分类实验表明,在多标记迁移学习中,基于局部标记关系的学习算法相比基于全局标记关系的学习算法具有更好的分类效果;本文所提算法与现有算法相比具有更好的分类效果.

Existing multi-label transfer learning algorithms employ the global label correlations. However, the global label correlations from source domain cannot be directly applied in the target domain due to the discrepancy between the source domain and target domain. In this paper, we propose a novel multi-label transfer learning algorithm based on local label correlations. The proposed algorithm can exploit the local label correlations shared between the source domain and the target domain by labelclustering and joint loss function minimizing. The corresponding local codes can be utilized as the auxiliary feature which can improve the learning model. Experimental results on multi-label imageclassification demonstrate that, the algorithms based on local label correlations deliver better than the algorithms based on global correlations in multi-label transfer learning problem, and the proposed algorithm delivers better performance than existing multi-label transfer learning algorithms.