Any-Cost Discovery: Learning Optimal classification Rules

Abstract

Fully taking into account the hints possibly hidden in the absent data, this paper proposes a new criterion when selecting attributes for splitting to build a decision tree for a given dataset. In our approach, it must pay a certain cost to obtain an attribute value and pay a cost if a prediction is error. We use different scales for the two kinds of cost instead of the same cost scale defined by previous works. We propose a new algorithm to build decision tree with null branch strategy to minimize the misclassification cost. When consumer offers finite resources, we can make the best use of the resources as well as optimal results obtained by the tree. We also consider discounts in test costs when groups of attributes are tested together. In addition, we also put forward advice about whether it is worthy of increasing resources or not. Our results can be readily applied to real-world diagnosis tasks, such as medical diagnosis where doctors must try to determine what tests should be performed for a patient to minimize the misclassification cost in certain resources.