Chapter 12 6) Conclusions and further analyses

This exercise allowed you to familiarize with Random Forest, by the proposed application about landslides susceptibility mapping and variables importance assessment.

To be sure that everything is perfectly clear for you, we propose you to answer the following questions and to discuss your answers with the other participants to the course or directly with the teacher.

1. Why is it important to implement the pseudo-absence, other that the presences (i.e., the observations) in a data-driven modelization?

2. What is the difference between a numerical and a categorical variable? Give some examples of both types. Why RF can handle directly native categorical variables? Is it the same for other algorithms (like logistic regression or neural network)?

3. Which is the values of the OOB estimate error rate of your model? Which parameters you can change to try to reduce it? Be brave and do it (i.e., change the values for ntree and mtry, than analyse which values for the AUC you obtain and which model perform better.