Data Mining and Predictive Analytics, 2nd Edition

Chapter 26Model Voting and Propensity Averaging

In Part 6: Enhancing Model Performance, we are examining methods for improving the performance of our classification and prediction models. In Chapter 24, we learned about segmentation models, where useful segments of the data are leveraged to enhance the overall effectiveness of the model. Then, in Chapter 25, we learned about ensemble methods, which combine the results from a set of classification models, in order to increase the accuracy and reduce the variability of the classification. Now, here in this chapter, we consider methods for combining different types of models, using model voting and propensity averaging.

26.1 Simple Model Voting

In Olympic figure skating, the champion skater is not decided by a single judge alone, but by a panel of judges. The preferences of the individual judges are aggregated using some combination function, which then decides the winner. In data analysis, different classification models (e.g., CART (classification and regression trees) vs logistic regression) may provide different classifications for the same data. Thus, data analysts may also be interested in combining classification models, using model voting or propensity averaging, so that the strengths and weaknesses of each model are smoothed out through combination with the other models. Model voting and propensity averaging are considered to be ensemble methods, because the ultimate classification decision is based, in part, on the input ...