On Exploring Soft Discretization of Continuous Attributes

Searching for a binary partition of attribute domains is an important task in data mining. It is present in both decision tree construction and discretization. The most important advantages of decision tree methods are compactness and clearness of knowledge representation as well as high accuracy of classification. Decision tree algorithms also have some drawbacks. In cases of large data tables, existing decision tree induction methods are often inefficient in both computation and description aspects. Another disadvantage of standard decision tree methods is their instability, i.e., small data deviations may require a significant reconstruction of the decision tree. We present novel soft discretization methods using soft cuts instead of traditional crisp (or sharp) cuts. This new concept makes it possible to generate more compact and stable decision trees with high accuracy of classification. We also present an efficient method for soft cut generation from large databases.