Hypercube processing of mixed sensed data entropic associations

摘要

A method for calculating unbiased entropic estimates of multivariate associations between mixed data is given. Since there is no assumption of unimodality of the distributions of the categorical and continuous-valued data, measures of central dispersion are not appropriate for the quantification of association. Empirical estimates of entropic associations are provided with respect to the partition entropy of a uniform binning interval and the cardinality of the sensed data. The increased computational demand incurred by the appropriate generalized measure is mitigated by a branch and bound algorithm for information-optimal attribute selection. The methodology is applied against a known data set used in a standard data mining competition that features both sparse categorical and continuous valued descriptors of a target with promising results.