Sparse Fisher's linear discriminant analysis for partially labeled data

• DOI: 10.1002/sam.11367
• Publication Date: 2017-12-06T13:08:59Z
• Authors: Qiyi Lu; Xingye Qiao
• Keywords: Linear classifier

Classification is an important tool with many useful applications. Fisher's linear discriminant analysis ( LDA ) a traditional model‐based classification method which makes use of the Gaussian distributional information. However, in high‐dimensional, low‐sample‐size setting, cannot be directly deployed because sample covariance not invertible. While there are modern methods for high‐dimensional data, they may fully information as does. Hence some situations, it still desirable to classification. This paper exploits potential more complicated data setting. In real applications, costly manually place labels on observations; consequently, often only small portion labeled available while large number observations left without labels. It great challenge obtain good performance through alone, especially order overcome this issue, we propose semisupervised sparse classifier take advantage seemingly useless unlabeled helps boost situations. A direct estimation used reconstruct and achieve sparsity; meanwhile employ difference‐convex algorithm handle nonconvex loss function associated data. Theoretical properties proposed studied. Our simulated examples help understand when how extracted from can useful. example further illustrates usefulness method.