Awards: Best Paper Honorable Mention Award at the SIAM (Society for Industrial and Applied Mathematics Conference on Data Mining (SDM). Paper: Scaling log-linear analysis to datasets with thousands of variables. Businesses and Government are investing heavily in their data assets. As data quantities continue to grow rapidly, it is increasingly difficult to extract maximum value from those data stores. Learning to predict the future from past observations is one of the key components that make it possible to bring value to data. To date, much of the research effort has been devoted to drawing predictions about a single pre-defined target variable, such as predicting the magnitude of the global warming, or the probability of developing cancer. However, in many real-world applications, what we wish to predict can change dramatically from one instance to the next, e.g. from one tactical situation to another or from one client to another. State-of-the-art techniques only provide ad-hoc solutions to this problem, because learning one model for every possibly encountered situation does not scale to big data assets. This project investigated methods for learning a single model that can effectively and efficiently predict all unobserved variables from the currently available evidence. New technologies were developed to learn models with this property from large and high dimensional data. The results show that the developed techniques offer a gain of 4 orders of magnitude in computation time over the state of the art.