Efficacious End User Measures—Part 1: Relative Class Size and End User Problem Domains

Figure 4

These graphs show that F-score, a summary statistic commonly used to compare CPD effectiveness, is rCS-sensitive. This is a desirable characteristic for problem domains such as information retrieval. In addition to rCS sensitivity, F-score is also sensitive to the target class.

(a) The relatively flat area around the optimum boundary (black dashed line in the graph) with low rCS suggests a low boundary sensitivity. The ridge follows the same optimum boundary as that of the total accuracy rate

(b) If, instead of selecting 𝑇₊, we select 𝑇₋, then F-score looks remarkably similar to TAR. Thus, F-score quantifies the categorical process’s effect on a specific category