Spare Projections with Pairwise Constraints

Abstract

In this paper, we propose a new semi-supervised DR method called sparse projections with pairwise constraints (SPPC). Unlike many existing techniques such as locality preserving projection (LPP) and semi-supervised DR (SSDR), where local or global information is preserved during the DR procedure, SPPC constructs a graph embedding model, which encodes the global and local geometrical structures in the data as well as the pairwise constraints. After obtaining the embedding results, sparse projections can be acquired by minimizing a L1 regularization-related objective function. Experiments on real-world data sets show that SPPC is superior to many established dimensionality reduction methods.