Measuring the correlation between variables based on the probability density function estimation

Abstract

Mutual information (MI) is always used as the indicator of nonlinear correlation between the variables. The computation of MI can be finished only the continuous-value variables are discretized. In this paper, one new strategy of computing the MI between variables is proposed. The probability density estimation (PDE) is used to determine the density functions in our method. An approximate technology is applied to replace the computation of integral. Finally, MI based on PDE can be obtained. Through the artificially experimental simulations, the performance and rationality of our new method are demonstrated. The experimental results show that our method is feasible, effective and efficient.