Spectral Partial Least Squares Regression

Linear Graph Embedding (LGE) is the linearization of graph embedding, and has been applied in many domains successfully. However, the high computational cost restricts these algorithms to be applied to large scale high dimensional data sets. One major limitation of such algorithms is that the generalized eigenvalue problem is computationally expensive to solve especially for large scale problems. Spectral regression can overcome this difficulty by casting the problem of learning an embedding function into a regression framework to avoid eigen-decomposition of dense matrices. In this paper, we develop a algorithm, Spectral Partial Least Squares Regression (SPLSR), which have advantages of PLSR and spectral regression. The experimental results have demonstrated the effectiveness of our proposed algorithm.