This arXiv preprint addresses semi-supervised learning for singular data sets. The authors propose to model such data clouds using subanalytic sets, a class of geometric sets that can describe non-smooth and irregular shapes.
The approach relies on a Morrey-type embedding for Sobolev spaces relative to globally subanalytic measures. This embedding is the key theoretical tool that lets the authors develop their semi-supervised learning method in this setting.
The abstract gives limited detail on the algorithm itself, focusing instead on the mathematical foundation. The work is primarily theoretical, offering a new way to think about learning on data with singular structure.