Description

Assignment 2
Refer to the Air Quality data described previously, and the analyses we have done with Ozone as the response variable, and the five explanatory variables (including the two engineered features from Assignment 1 as below).
• AQ$TWcp = AQ$Temp*AQ$Wind
• AQ$TWrat = AQ$Temp/AQ$Wind
1. Fit a default Random Forest (RF) to only the three main variables in the data—Temp, Wind, and Solar.R—and not the two extra ones that we engineered. A RF should be able to detect interactions automatically if needed.
(a) Report the OOB error.
(b) Produce variable importance measures and comment on the relative importance of the variables. How do they compare to what we have seen in earlier analyses of these data?
2. Repeat the exercise in question 1, adding the two engineered features into the data.
(a) Report the OOB error. Does it improve much compared to the previous RF analysis without the variables?
(b) Produce variable importance measures. Are the two engineered features particularly important?
3. Use boosting to model the relationship between Ozone and all ONLY THE THREE ORIGINAL VARIABLES. Tune on an initial grid of λ = 0.001,0.005,0.025,0.125 and d = 2,4,6, and select trees optimally using twice the number suggested by OOB error. Use two reps of 5-fold CV (refer to the lecture note and R code to understand how to do this).
(a) Report the mean root-MSPE for each combination of λ and d (b) Show relative root-MSPE boxplots for each combination of λ and d (c) What combination of λ and d do you prefer?
1