Classification of discrete data with feature space transformation

A newly developed classification scheme for samples with discrete valued features is presented in this paper. In it, we first map the discrete feature space into a Euclidean space called logarithm of likelihood ratio (LLR) space. The likelihood ratios are formed from the estimated distributions based on the dependence tree structure obtained through minimizing the error probability. By discriminant analysis, we then transform the LLR space into one-dimensional space on which classification is conducted. We have applied this new scheme to several sets of biomedical data and have obtained significantly high classification rates.