New methods for finding associations in large data sets: Generalizing the maximal information coefficient (MIC)

We propose here a natural, but substantive, extension of the MIC. Defined for two variables, MIC has a distinct advance for detecting potentially complex dependencies. Our extension provides a similar means for dependencies among three variables. This itself is an important step for practical applications. We show that by merging two concepts, the interaction information, which is a generalization of the mutual information to three variables, and the normalized information distance, which measures informational sharing between two variables, we can extend the fundamental idea of MIC. Our results also exhibit some attractive properties that should be useful for practical applications in data analysis. Finally, the conceptual and mathematical framework presented here can be used to generalize the idea of MIC to the multi-variable case.