Recursive Orthogonal Label Regression: A Framework for Semisupervised Dimension Reduction

Semisupervised DR techniques using virtual label regression have attracted considerable attention, but they suffer from two restrictions: the number of discriminant directions available is constrained to the number of classes, and they're nonorthogonal. Traditional methods easily address these problems. However, an interesting problem is how to address these restrictions in label regression modelings. To do this, the authors developed Recursive Orthogonal Label Regression (ROLR), a regression framework of semisupervised dimension reduction that uses label propagation and label regression in a recursive procedure. Here, they illustrate the formulation of ROLR using semisupervised regression encoding. ROLR provides an unified view to understand and explain a large family of label regression techniques. Experimental results show the approach's feasibility and effectiveness.