Evaluating Classifiers
for Changing Environments
ROC curves and the ROC Convex Hull
Why not just use classification accuracy?
Why not just use classification accuracy?
We could calculate expected cost/profit
Unfortunately, target conditions are difficult to know exactly . . .
ROC space and iso-performance lines
Different conditions identify different “best” classifiers
Using the ROCCH as an optimal hybrid classifier
Summary
ROC Convex Hull
(ROCCH)
[End]
supplementary
slides follow
Advanced Topics
Variable
costs/benefits (case by case)
Advanced Topics
Variable
costs/benefits (continued)
Maximizing precision at cutoff
Precision/recall/coverage curves