We can imagine several ways to implement this. For instance, one can perform a k-NN estimation of the label. Another option would be to use PAk interpolation and then follow the rule of assigning it to the nearest point with a higher density. However, these procedures should be tested and are not part of the original clustering algorithm. I don't think this feature is critical at the moment, we can leave it as in the case of spectral clustering on sklearn, in which the predict option is not there and only the fit and fit_predict work. https://scikit-learn.org/stable/modules/generated/sklearn.cluster.SpectralClustering.html#sklearn.cluster.SpectralClustering