Mining Strongly Correlated Intervals with Hypergraphs

Correlation is an important statistical measure for estimating dependencies between numerical attributes in multivariate datasets. Previous correlation discovery algorithms mostly dedicate to find piecewise correlations between the attributes. Other research efforts, such as correlation preserving discretization, can find strongly correlated intervals through a discretization process while preserving correlation. However, discretization based methods suffer from some fundamental problems, such as information loss and crisp boundary. In this paper, we propose a novel method to discover strongly correlated intervals from numerical datasets without using discretization. We propose a hypergraph model to capture the underlying correlation structure in multivariate numerical data and a corresponding algorithm to discover strongly correlated intervals from the hypergraph model. Strongly correlated intervals can be found even when the corresponding attributes are less or not correlated. Experiment results from a health social network dataset show the effectiveness of our algorithm.