基于半监督假设的半监督稀疏度量学习

传统的有监督度量学习算法没有利用大量存在的无标记样本,且得到的度量矩阵复杂,难以了解不同原始特征的重要程度。针对这些情况,提出基于半监督假设的半监督稀疏度量学习算法。根据三样本组约束建立间隔损失函数;基于平滑假设、聚类假设、流形假设这三个半监督假设建立半监督正则项,并利用L1范数建立稀疏正则项;利用梯度下降法求解目标函数。实验结果表明,该算法学习得到的度量能有效地使不同类别的样本间距离增大,度量矩阵具有稀疏性,分界面穿过低密度区域,该算法在UCI的样本数据集上具有良好的分类准确性。

SEMI-SUPERVISED SPARSE METRIC LEARNING BASED ON SEMI-SUPERVISED ASSUMPTION

Traditional supervised metric learning algorithms do not take advantage of the large number of unlabeled samples,and the resulting metric matrix is complex.So it is difficult to understand the importance of different input features.To solve these problems,we proposed a semi-supervised sparse metric learning algorithm based on semi-supervised assumption.The interval loss function was established according to the triplet constraints.Then,we established a semi-supervised regularization based on the three semi-supervised assumptions: smoothness assumption,cluster assumption and manifold assumption,and established a sparse regularization term by using L 1 norm.The gradient descent method was used to solve the objective function.The experimental results show that the metrics learned by the algorithm can effectively increase the distance between different types of samples,and the metrics matrix is sparse.The interface crosses the low-density region.The algorithm has good classification accuracy on the UCI sample.