Bank Product Marketing This problem uses the dataset bank-bias-data.csv. 13 A bank with operations in the United States and Europe is seeking to promote a new term deposit product with a direct marketing campaign to its customers (primarily via telemarketing). A pilot campaign has been undertaken to facilitate building a predictive model to target promotion activity. Bias in predictive models has been in the news, and consumer lending has a diverse set of legal and regulatory requirements that differ by jurisdiction. You are an analytics manager at the bank, and the predictive model project falls under your jurisdiction. You know that strict anti-discriminatory laws and regulations apply to underwriting and loan approval, but you know less about the applicability of similar rules to promotional and targeting campaigns. Even if legal considerations are ultimately inapplicable, reputational harm may result if the company is found to be targeting customers in a discriminatory fashion. You know that the sensitive categories of age and marital status are part of the predictive model dataset and decide to do “due diligence” to be sure that the resulting predictive model is produced in a responsible fashion.

a. Explore the data to assess any imbalance in the target attribute “deposit,” across sensitive categories (age and marital status).

b. Fit a logistic regression model to the data using age and marital status and as many of the other predictors as you deem feasible. Report the regression equation.

c. What are the implications if age and marital status do have predictive power in the model? What are the implications if they do not?

d. Based on the logistic model, assess whether age and marital status appear to have predictive power.

e. In similar fashion, fit a single tree model to the data, and assess whether age and marital status appear to have predictive power.

f. Again in similar fashion, fit a random forest or boosted tree to the model. Use a permutation feature importance test to assess whether age and marital status have predictive power.

g. Compare the three models’ performance in terms of lift, and sum up your findings about the roles played by the sensitive predictors.