There is considerable interest in developing techniques for predicting human behavior, for instance to enable emerging contentious situations to be anticipated or permit the nature of ongoing but “hidden” activities to be inferred. A promising approach to this problem is to collect appropriate empirical data and then apply machine learning methods to the data to generate the predictions. This two-part paper shows that the performance of such learning algorithms often can be improved substantially by leveraging sociological models in their development and implementation. In particular, we demonstrate that sociologically-grounded learning algorithms outperform gold-standard methods in two important and challenging tasks: 1.) inferring the (unobserved) nature of relationships in adversarial social networks, and 2.) predicting whether nascent social diffusion events will “go viral”. Significantly, the new algorithms perform well even when there is limited data available for their training and execution.