1: \begin{abstract}

2: We describe a new framework for graph-based semi-supervised learning. Motivated by the need to address the degeneracy of canonical Laplace learning algorithms in low label rates, we reformulate graph-based semi-supervised learning as a generalization of a \emph{Trust-Region Subproblem} (TRS) in which one is asked to optimize a nonconvex quadratic over a Euclidean sphere. Our method is motivated by the well-posedness of Laplacian eigenvectors in the limit of infinite unlabeled data. 

3: %as well as recent developments regarding the efficacy of normalization on finding solutions to Laplace equations.

4: %Despite the nonconvexity of typical TRS problems, 

5: We propose approximate and iterative algorithms that enjoy global convergence guarantees. We first show that a first-order condition implies the solution of a manifold alignment problem and that solutions to the classical \emph{Orthogonal Procrustes} problem can be used to find good initial solutions. Next, we describe an efficient subspace method and prove that high-quality local solutions lie in a subspace that includes the span of principal eigenvectors of the graph Laplacian. We demonstrate that our method achieves lower classification error compared to recent state-of-the-art and classical semi-supervised learning methods at extremely low, medium, and high label rates. Our code is available on github\footnote{anonymized for submission}.

6: 

7: \end{abstract}

8: