Feature ordering and stopping rule based on maximizing mutual information

Summary form only given. The problem of feature ordering and stopping rule for sequential Bayesian classification is considered. A criterion that maximizes mutual information has been developed and compared with conventional strategies. At each stage the feature that maximizes mutual information gain from the observed data is selected, and the sequential procedure is terminated if its maximum value is less than a positive constant C. The advantages of the scheme are outlined. Numerical results have shown the good behavior of the proposed technique if the number of patterns or the allowable average number of used features is large. It has been shown that this scheme reduces the misallocation error rate, especially in the early stage, with the same mean number of used features.