Feature selection and graphical model structure learning

Feature selection is a technique used in machine learning (classification or regression) to select a subset of the input features so that the problem size is decreased while the learning performance is not compromised. Graphical model structure learning is to learn the graph structure that best represents the independency relations among the random variables of the target probabilistic distribution. At first glance, the two don't seem to be related. But think about it, feature selection tries to find the minimal set of variables that are sufficient to predict the target variable, which is the Markov blanket of the target variable; and finding the Markov blanket is obviously helpful in learning the graphical model structure (and the reverse is also true). This is exactly what this paper "Using Markov Blankets for Causal Structure Learning" in a recent issue of JMLR does.

Btw, the first two posts of this blog are both about learning graphical model structure, but... that's mere coincidence! My current research is only remotely related to this area. I'll talk about some other topic in the next post.