基于GPU并行加速的多特征融合的超图降维方法  

Hypergraph Dimensionality Reduction with Multiple Feature Fusion Based on GPU Parallel Acceleration

作　　者：洪朝群[1] 陈旭辉[1] 王晓栋[1] 李士锦[1] 吴克寿[1] 

机构地区：[1]厦门理工学院计算机与信息工程学院,厦门361024

出　　处：《计算机科学》2015年第11期90-93,117,共5页Computer Science

基　　金：国家自然科学基金(61202145);福建省自然科学基金(2014J01256)资助

摘　　要：基于图的学习方法目前广泛用于降低特征维度。然而,对于多特征数据而言,不同特征之间的不同关联性很难结合到单个图中。针对多特征数据提出了新的半监督降维方法。首先,以超图中的超边作为片,使超图应用到片对齐框架中。然后,通过统计片中相邻的特征对的距离计算超边的权重,使得不同特征下的片得到结合。其次,由于欧氏距离和矩阵乘法的计算在拉普拉斯矩阵的构造过程中占用了大部分的时间,因此使用GPU对其进行加速。实验结果表明了所提方法在分类性能和学习速度上的提升效果。Graph-based learning methods are currently popular for dimensionality reduction. However, for multiple fea- ture data, different relationships from different features are hard to be integrated into a single graph. In this paper,a no- vel semi-supervised dimensionality reduction method was proposed for multiple feature data. First, the hyperedges in hy- pergraph are assumed as patches. In this way, hypergraph is applied to patch alignment framework. Then, the weights of hyperedges are computed with statistics of distances between neighboring pairs and the patches from different features are integrated. Second, the speed of computing Euclidean distances and matrix multiplication is improved by using GPU, since they take most of time in constructing the Laplacian matrix. The experimental results demonstrate the improve- ment on both classification performance and learning speed.

关 键 词：降维 多特征融合 片对齐框架 超图学习 基于GPU的并行加速 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]