One of the most pervasive metaphors in TDA is that homology generalizes clustering. This situates techniques like persistent homology firmly in the domain of unsupervised learning. Although less common, there is also substantial work on the estimation of a persistence barcode of an unknown function from samples---a kind of supervised TDA problem. In this talk, I will consider the question of what kinds of guarantees are possible if one has evaluations of the function at only a subset of the input points---a semi-supervised TDA problem. I will summarize the new theory of sub-barcodes and show how one can compute a barcode that is guaranteed to be contained in the barcode of every Lipschitz function that agrees with the sample data. This provides strong guarantees even with only partial data.