Graphs play a role in many semi-supervised learning algorithms, where unlabeled samples are used to find useful structural properties in the data. Dimensionality reduction and regularization based on preserving smoothness over a graph are common in these settings, and they perform particularly well if proximity in the original feature space closely reflects similarity in the classification problem of interest. However, many real-world problem spaces are overwhelmed by noise in the form of features that have no useful relevance to the concept that is being learned. This leads to a lack of robustness in these methods that limits their applicability to new domains. We present a graph-construction method that uses a collection of task-specific random subspaces to promote smoothness with respect to the problem of interest. Application of this method in a graph-based semi-supervised setting demonstrates improvements in both the effectiveness and robustness of the learning algorithms in noisy problem domains.