gistfile1.txt

m_gbm_0 = train(x=df_0, y=labels_0, 
              method="gbm", weights = weight_0,
              verbose=TRUE, trControl=ctrl, metric="AMS")
m_gbm_1 = train(x=df_1, y=labels_1, 
                method="gbm", weights=weight_1, 
                verbose=TRUE, trControl=ctrl, metric="AMS")
m_gbm_2 = train(x=df_2, y=labels_2, 
                method="gbm", weights=weight_2, 
                verbose=TRUE, trControl=ctrl, metric="AMS")
m_gbm_3 = train(x=df_3, y=labels_3, 
                method="gbm", weights=weight_3, 
                verbose=TRUE, trControl=ctrl, metric="AMS")

gbmTrainPred_0 <- predict(m_gbm_0, newdata=df_0, type="prob")
gbmTrainPred_1 <- predict(m_gbm_1, newdata=df_1, type="prob")
gbmTrainPred_2 <- predict(m_gbm_2, newdata=df_2, type="prob")
gbmTrainPred_3 <- predict(m_gbm_3, newdata=df_3, type="prob")


After this, we determined the ideal threshold over which to predict our gbm model. The ideal thresholds were different for each data frame, and we accounted for these accordingly. An example output plot for deterimining the threshold can be seen below.

auc_0 = roc(labels_0_n, gbmTrainPred_0[,2])
auc_1 = roc(labels_1_n, gbmTrainPred_1[,2])
auc_2 = roc(labels_2_n, gbmTrainPred_2[,2])
auc_3 = roc(labels_3_n, gbmTrainPred_3[,2])

plot(auc_0, print.thres=TRUE)
plot(auc_1, print.thres=TRUE)
plot(auc_2, print.thres=TRUE)
plot(auc_3, print.thres=TRUE)


threshold_0 <- 0.001
threshold_1 <- 0.002
threshold_2 <- 0.006
threshold_3 <- 0.005