Training Data [4] 50% 2421/4843 4843(0.05) = 242.15 ___|_____ 44% [5] [2] 1216/2112 = 57.6% ____|___ 42% [6] [1] 115/188 = 61.2% ___|____ [3] [7] 836/2071 = 40.4% 53.8% 254/472 0 188 61.2% 115 188 | | 5% 10% 242.15 2300 57.6% 1216 2112 … 2772 53.8% 254 472 | | 45% 50% 2179.35 2421.5 4843 40.1% 836 2071 | | 55% 60% More Evaluations: Decisions: Y = Sensitivity: Probability of calling a 1 a 1. Pr{decide 1 |1}.Want large Y. Specificity: Probability of not calling a 0 a 1. X = 1-Specificity: Probability of mistakenly calling a 0 a 1. Pr{Decide 1 | 0}. Want small X. ROC Plot (receiver operating characteristic curve) Declare all data 0 Sensitivity=0, 1-Specificity=0 (0,0) ------------------------------------------------------------------------------------------Cut at leaf 1: Leaf 1 data declared 1’s (115 right, 188-115=73 wrong <– 0 called 1) Sensitivity: 115/2421 = 0.0475 1-Specificity: 73/2422 = 0.0301 (0.0301, 0.0475) ------------------------------------------------------------------------------------------Cut at leaf 2: Leaf 1 and 2 declared 1’s (1331 right, 2300-1331= 969 wrong <- 0 called 1) Sensitivity: 1331/2421 = 0.5498 1-Specificity: 969/2422 = 0.4000 (0.4000, 0.5498) -------------------------------------------------------------------------------------------Cut at leaf 3: Sensitivity: 1585/2421 = 0.6547 1-Specificity: 1187/2422 = 0.4900 (0.4900 , 0.6547) -------------------------------------------------------------------------------------------Declare all 1’s: Sensitivity = 1, 1-Specificity = 1 (1,1) Plot these 5 points, connect the dots – that’s the ROC curve! You want to maximize area under ROC curve (maximizes the concordance) You can plot ROC for a set of models even if one is regression, another a tree, etc. Percent Captured Response: There are 2421 1’s. First 5% captures an estimated 146.177 of them: 146.177/2421 = 6.0379% Next 5% captures an estimated 242.15(1216/2112)= 139.4197 of them: 5.7587% Next few 5% sets capture 5.7587% as well (all in leaf 2) SAS gives cumulative % captured 1’s 6.0379%, 6.0379%+5.7587%, 6.0379%+2(5.7587%), etc. Ideal cumulative percent response, lift, etc. : If you had a perfect predictor, it would correctly call all the 1’s and 0’s so the cumulative percent response moving across the “deciles”, for example, would be cumulative percent response = 100% until it got to the 50thpercentile (in our case) where it would start picking up 0’s and decreasing from 100% towards 50%, hitting that overall 50% response rate in our data at the 100th percentile.