The “adult” data set at the UCI Machine Learning Repository is derived from census records.5 In these data, the goal is to predict whether a person’s income was large (defined in 1994 as more than $50K) or small. The predictors include educational level, type of job (e.g., never worked, and local government), capital gains/losses, work hours per week, native country, and so on.6 After filtering out data where the outcome class is unknown, there were 48842 records remaining. The majority of the data were associated with a small income level (75.9 %). The data are contained in the arules package and the appropriate version can be loaded using data (AdultUCI).

(a) Load the data and investigate the predictors in terms of their distributions and potential correlations.

(b) Determine an appropriate split of the data.

(c) Build several classification models for these data. Do the results favor the small income class?

(d) Is there a good trade-off that can be made between the sensitivity and specificity?

(e) Use sampling methods to improve the model fit.

(f) Do cost-sensitive models help performance?