Scalable Laplacian Regularized Least Squares Classification on Anchor Graph

In this paper, we address the scalability issue of graph-based semi-supervised learnings. In recent years, graph-based semi-supervised learnings using anchor graphs and hierarchical anchor graphs have been proposed as high scalability approaches. However, these techniques utilize linear predictor and cannot infer to unknown data. We propose a Laplacian regularized least squares classification to assume a hierarchical anchor graph (LapRLSC-HAG). The proposed method reduces a memory pressure by the property of the anchor graph or hierarchical anchor graph. Furthermore, we present two strategies optimizing the primal LapRLSC-HAG problems to reduce the computational complexity. To demonstrate the effectiveness of the proposed method, we make experimental evaluations on data sets varying the size from thousands to one million.