A two-tiered 2D visual tool for assessing classifier performance.

In this article, a new kind of 2D tool is proposed, namely ⟨φ, δ ⟩ diagrams, able to highlight most of the information deemed relevant for classifier building and assessment. In particular, accuracy, bias and break-even points are immediately evident therein. These diagrams come in two different forms: the first is aimed at representing the phenomenon under investigation in a space where the imbalance between negative and positive samples is not taken into account, the second (which is a generalization of the first) is able to visualize relevant information in a space that accounts also for the imbalance. According to a specific design choice, all properties found in the first space hold also in the second. The combined use of φ and δ can give important information to researchers involved in the activity of building intelligent systems, in particular for classifier performance assessment and feature ranking/selection. [ABSTRACT FROM AUTHOR]