The cluster assumption is the key ingredient underlying a large class of algorithms for graph-based Semi-Supervised Learning. In this paper we take the cluster assumption literally and use multiple clusterings obtained by 1-spectral clustering to define a data-adaptive basis of functions. This data-adaptive basis implements the cluster assumption and thus simple ridge regression can be used for Semi-Supervised Learning. We show that the aggregation of information from multiple clusterings leads to significantly better results than just using the best clustering and provide an interpretation in terms of label propagation. Furthermore, our approach leads to the best reported error rates for MNIST and USPS datasets with just one label per class.