A nominal association matrix with feature selection for categorical data

We introduce an informative probabilistic association matrix to measure a proportional local-to-global association of categories of one variable with another categorical variable. Towards a probability based proportional prediction, the association matrix gives rise to the expected predictive distribution of the first and second types of errors for a multinomial response variable. In addition, the normalization of the diagonal of the matrix gives rise to an association vector, which provides the expected category accuracy lift rate distribution. A general scheme of global-to-global association measures with flexible weight vectors is further developed from the matrix. A hierarchy of equivalence relations defined by the association matrix and vector is shown. Applications to financial and survey data together with simulations results are presented.