Abstract: A method for determining accuracy of a classifier which provides an indication of the degree of correctness of the classifier rather than a mere correct/incorrect indication. The accuracy of the classifier is determined by determining a set of categories of an arrangement of categories selected for an item by the classifier and determining a set of categories of the arrangement selected for the item by an authoritative classifier. An accuracy measure which indicates a degree of correctness of the classifier is then determined based on the categories selected by the classifier and the categories selected by the authoritative classifier.

Abstract: A method for determining accuracy of a classifier which provides an indication of the degree of correctness of the classifier rather than a mere correct/incorrect indication. The accuracy of the classifier is determined by determining a set of categories of an arrangement of categories selected for an item by the classifier and determining a set of categories of the arrangement selected for the item by an authoritative classifier. An accuracy measure which indicates a degree of correctness of the classifier is then determined based on the categories selected by the classifier and the categories selected by the authoritative classifier.

Abstract: A technique for automatically designing a fraud detection system using a series of machine learning methods. Data mining and constructive induction are combined with more standard machine learning techniques to design methods for detecting fraudulent usage based on profiling customer behavior. Specifically, a rule-learning is used to uncover indicators of fraudulent behavior from a large user database. These indicators are used to create profilers, which then serve as features to the fraud detection system that combines evidence from multiple profilers to generate high-confidence intervention activities when the system is deployed on-line with user data.