加权成对约束半监督局部维数约减算法

Weight pairwise constraints based semi-supervised locality dimensionality reduction

考虑到已有的半监督维数约减方法在利用边信息时将所有边信息等同,不能充分挖掘边所含信息,提出加权成对约束半监督局部维数约减算法(WSLDR)。通过构建近邻图对边信息进行扩充,使边信息数量有所增加。另外,根据边所含信息量的不同构建边的权系数矩阵。将边信息融入近邻图对其进行修正,对修正后的近邻图和加权的成对约束寻找最优投影。算法不仅保持了数据的内在局部几何结构,而且使得类内数据分布更加紧密,类间数据分布更加分散。在UCI数据集上的实验结果验证了该算法的有效性。

A novel algorithm called weight pairwise constraints based semi-supervised locality dimensionality reduction （WSLDR） is proposed to overcome the shortcoming of the semi-supervised dimensionality reduction algorithms can not fully utilize the side information. First, the side information is expanded by k-nearest neighborhood graph, which may increase the number of the constraints, then weights are assigned to the side by its information power. Second, the side information is integrated into the neighbor graph to revise it. Last, a projection is found based on the revised neighborhood graph and the weighted constraints. The projects not only maintain the local geometric structure, but also decrease the distance within class, and increase the distance between classes. The experimental results on UCI datasets show the effectiveness of the algorithm.