To date, most network modeling has been descriptive, but a mature science requires predictive models whose predictions are empirically verified. It also requires a mathematical language in which these models are easily and compactly defined, and an efficient computational implementation of it. Markov logic is a language for network science, based on combining Markov networks and first-order logic. Here we extend Markov logic to handle decision-theoretic problems, and apply it to viral marketing. This has two components: learning a model of the social network from data, and using it to design a marketing plan that maximizes utility (profits minus marketing costs). Applied to the Epinions Web of Trust, our approach greatly outperforms previous ones. Our learning and inference algorithms are available in the open-source Alchemy system (http://alchemy.cs.washington.edu/).
(Joint work with Aniruddh Nath and Matt Richardson.)