Methods for discovery of local causes/effects and of Markov blankets enable discovery of causal relationships from large data sets and provide principled solutions to the variable/feature selection problem, an integral part of predictive modeling. The present invention provides a generative method for learning local causal structure around target variables of interest in the form of direct causes/effects and Markov blankets applicable to very large real world datasets even with small samples. The selected feature sets can be used for causal discovery, classification, and regression. The generative method GLL can be instantiated in many ways giving rise to novel method variants. The method transforms a dataset with many variables into either a minimal reduced dataset where all variables are needed for optimal prediction of the response variable, or a dataset where all variables are direct causes and direct effects or the Markov blanket of the response variable.