Chapter 13 – Ensembles and Uplift Data Mining for Business Analytics Shmueli, Patel & Bruce
Ensembles can improve predictive power Model 1 Model 2 Error e1,i e2,i Expected Error E(e1,i) = 0 E(e2,i) = 0 Variance Var(e1,i) Var(e2,i) Ensemble: E((e1,i+ e2,i)/2) = 0 ¼ Var(e1,i) + ¼ Var(e2,i) + Cov(e1,i, e1,i)
Methods used Simple averaging Weighted averaging Voting for classifiers
Bagging Bootstrap sampling Aggregating Generate multiple random samples with replacement Aggregating Run modeling algorithm on each sample Combine the results
Boosting Fit a model Generate a sample -- oversample misclassified cases Fit the model to the new sample Repeat 2 and 3 multiple times Bagging improves stability Helps avoid over fitting
Advantages and Disadvantages More precise predictions Improves stability Helps avoid over fitting Disadvantages Requires more resources Time Ensemble model non-interpretable
Uplift modeling Collect sample data including current status Randomly split data into treatment and control group Apply treatment to treatment group Measure status change for both groups Recombine sample and randomly partition into training and validation sets Develop model to training set with Status change as target and include treatment applied or not as a predictor
Uplift modeling ….. Run model to validation data set with treatment set to 1 and calculate propensity, i.e. P(Success/Treatment = 1) Repeat with treatment set to 0 and calculate propensity, i.e. P(Success/Treatment = 0) Uplift = P(Success/Treatment = 1) - P(Success/Treatment = 0)
Uplift Example
Uplift Example
Uplift Example
Uplift Example