监督型稀疏保持投影

Supervised sparsity preserving projections

稀疏保持投影(SPP)是最近提出的一种无监督降维方法,因此无法利用标号数据提供的监督信息。为此,对SPP进行了扩展,给出了两种监督型稀疏保持投影算法:基于稀疏保持的判别分析(SPP+LDA)和监督稀疏保持投影(S2PP)。前者通过在SPP变换的子空间内进行线性判别分析(LDA)达到利用数据间稀疏重建关系和监督信息的目的;后者借助数据标号直接修正SPP构建的稀疏重建图在SPP中自然地融入监督信息。分析了两种算法的优缺点,在两个常用的人脸数据集(Yale和AR)上验证了两者的可行性及有效性。

Sparsity Preserving Projection（SPP） is a recently proposed unsupervised dimensionality reduction method, thus fails to use the supervised information provided by the labeled data.To address this issue,two supervised algorithms for extending SPP are presented, called SPP-based Linear Discriminant Analysis（SPP＋LDA） and supervised SPP（S2PP） respectively. The former takes advantage of sparse reconstructive relationship and label information in data by applying LDA in the SPP transformed subspace, and the latter naturally incorporates discrimination information by utilizing label information to modify sparse reconstructive graph constructed via SPP.The advantages and disadvantages of the two proposed methods are analyzed. The feasibility and effectiveness of the proposed methods are verified on two popular face databases（Yale and AR） with promising results.