Search

Decision rules in if-then form are highly readable and suitable for the situations in which users need to understand the rules intuitively. When we suppose the situation in which someone reads rules, a set of decision rules is desired to satisfy the following three conditions: 1) They can explain most of possible situations as a rule set, 2) The size of a rule set is small and thus memorable, 3) Description of each rule is simple and easily understood. In general, however, it is difficult to achieve both 2) and 3) under the condition 1). In addition to typical reduction of attributes, we consider reduction of attribute domains, the number of possible attribute values in each attribute, aiming at obtaining simpler but more readable rules. It brings a large variety of granularity in data representation. Using previously proposed some criteria on the basis of 1) through 3), we rated rule sets obtained at specified levels of granularity in some real-life datasets. The rating was almost consistent to that by a human inspector in readability.

Tree-type expression of multi-class problems is known to be useful for drawing several insights from a given problem and for improving the performance of classifiers. The authors have already proposed a bottom-up procedure to construct such a tree, called a “class decision tree”, but a top-down procedure is still worth studying. In this paper, we propose a simple top-down procedure, and compare those two procedures. In addition, we discuss the effectiveness of classifier selection and feature selection applied to every node in class decision trees, and the preferable order of them as well.

Feature selection is an important technique in pattern recognition. By removing features that have little or no discriminative information, it is possible to improve the predictive performance of classifiers and to reduce the measuring cost of features. In general, feature selection algorithms choose a common feature subset useful for all classes. However, in general, the most contributory feature subsets vary depending on classes relatively to the other classes. In this study, we propose a classifier as a decision tree in which each leaf corresponds to one class and an internal node classifies a sample to one of two class subsets. We also discuss classifier selection in each node.

In pattern recognition, feature selection is an important technique for reducing the measurement cost of features or for improving the performance of classifiers, or both. Removal of features with no discriminative information is effective for improving the precision of estimated parameters of parametric classifiers. Many feature selection algorithms choose a feature subset that is useful for all classes in common. However, the best feature subset for separating one group of classes from another may depend on groups. In this study, we investigate the effectiveness of choosing feature subsets depending on groups of classes (class-dependent features), and propose a classifier system that is built as a decision tree in which nodes have class-dependent feature subsets.