From Distance Correlation to Multiscale Generalized Correlation
Understanding and developing a correlation measure that can detect general dependencies is not only imperative to statistics and machine learning, but also crucial to general scientific discovery in the big data age. We proposed the Multiscale Generalized Correlation (MGC) in Shen et al. 2017 as a novel correlation measure, which worked well empirically and helped a number of real data discoveries. But there is a wide gap with respect to the theoretical side, e.g., the population statistic, the convergence from sample to population, how well does the algorithmic Sample MGC perform, etc. To better understand its underlying mechanism, in this paper we formalize the population version of local distance correlations, MGC, and the optimal local scale between the underlying random variables, by utilizing the characteristic functions and incorporating the nearest-neighbor machinery. The population version enables a seamless connection with, and significant improvement to, the algorithmic Sample MGC, both theoretically and in practice, which further allows a number of desirable asymptotic and finite-sample properties to be proved and explored for MGC. The advantages of MGC are further illustrated via a comprehensive set of simulations with linear, nonlinear, univariate, multivariate, and noisy dependencies, where it loses almost no power against monotone dependencies while achieving superior performance against general dependencies.
READ FULL TEXT