Speedy local search for semi-supervised regularized least-squares

In real-world machine learning scenarios, labeled data is often rare while unlabeled data can be obtained easily. Semi-supervised approaches aim at improving the prediction performance by taking both the labeled as well as the unlabeled part of the data into account. In particular, semi-supervised support vector machines favor decision hyperplanes which lie in a "low-density area" induced by the unlabeled patterns (while still considering the labeled part of the data). The associated optimization problem, however, is of combinatorial nature and, hence, difficult to solve. In this work, we present an efficient implementation of a simple local search strategy that is based on matrix updates of the intermediate candidate solutions. Our experiments on both artificial and real-world data sets indicate that the approach can successfully incorporate unlabeled data in an efficient manner.