Data Modelling/Statistical Learning - Interview Questions

Question

I was asked the following 2 questions in an interview. I wasn't selected which means my answers were wrong.

Now I need to learn from my mistakes. I have though quite a bit since the interview and still not got anywhere concrete so need your help..

Question 1 : It is raining during evening and you cant go out. The TV news broadcast says that it will continue to rain for next 5 hours. What are the chances that you will see the sun in the next 36 hours ?

My Answer in Interview : I said something to the effect that I will look at the data throughout the month and past years(same month) to see on average how often we get the clear skies after rainy day.

I think I should have said that this is not a viable scenario to apply statistical learning

Question 2 : In a company which had 1000 employees, 100 employees left in a year. Now the company has asked you to look at the existing employees and find out who are at the risk of leaving. The company can provide you any data that you would want.

My answer : I said I would look at the salaries of these people and those who left. Compare it with market standards and see if I can find any pattern. Also I said it would look at the ratings of the people leaving and see if they are star performers or not, since high performers who do not get promoted may also leave. I would use logistic regression for classification.

I felt I was on the right track but obviously that wasn't the case.

I feel the interviewer was looking for my basic approach rather than any complicated statistical learning methods.

Answer 1

For the second question, I don't think people want to know specifically which model that you going to use because it is normally decided after some cross-validation. In this case, salary is a good feature to be included but there might also be: years of experience, department, age, degree, gender, marital status, etc. The provided data includes 100 positive and 900 negative labels, so I would tell how I deal with this unbalanced data: either under-sampling or under weighting those 900 negative-labelled samples. Finally, I would mention cross-validation to choose the best model, the best strategy to deal with the unbalanced data, and maybe the best feature set if there is too many features available. Besides, since we are interested in the risk of leaving, I might want to use Area Under Curve (AUC) instead of simple error rate as evaluation metric.

For the first question, since you watched the weather forecast, you should know the answer. Building a model is OK but it is too much harder and often results in a less accurate prediction considering that you have less resources than those who do it for the TV broadcast. Looking for historical weather data is much harder than for a 48-hours weather forecast!