Important statistical considerations in classifier systems

The performance of a classifier system may be limited due to the following: (1) nonmonotonic relationships between individual predictor co-factors and outcomes, (2) prevalence imbalances between development data and application environment data, and (3) failure to account for cost-gain economics. These issues are explored, and statistically-based techniques for treating them are presented. In addition, probabilistic and fuzzy interpretations of classifier outputs are discussed, a likelihood ratio transformation of classifier outputs is suggested and two new cost-gain indexes that rate classifier systems in global economic terms are introduced.