9: Uplift Modeling and Causal Machine Learning

Question 1

Question 1
Level: medium
When collecting data for developing a churn prediction model over a time window during which selected customers are targeted with a customer retention campaign, then which of the following labels can NOT apply?

a) Persuadables may be labeled as non-churners.
b) Lost causes may be labeled as non-churners.
c) Sure things may be labeled as non-churners.
d) Do-not-disturbs may be labeled as churners.

Answer 1

b) Lost causes may be labeled as non-churners.

x-axis; churn when targeted
y-axis; churn when not targeted
values;
Persuadable, lost cause
Sure things, sleeping dogs

Lost causes churn regarless of action,
Sure things dont churn
Persuadables are bout to churn, u have to stop them
Sleeping dogs, are asleep and wont churn if you dont target them, will churn if you target them

Question 2

Question 1
What are hidden confounders and why are these of importance to causal effect learning?

Answer 2

2. Challenges in applying ITE modeling (Individual treatment effect)
   - Inversed causality; when cause-and-effect relationship between variables is revrsed.
   - Hidden confounder; unobserved variables that can affect both the treatment and outcome, leading to biased estimates. Crucial to address this.
   Wrong conclusions w.r.t. causal relationships in causal discovery
   Wrong models and estimates of treatment effects in causal effect learning -> solve via control vs treatment group aka randomized controlled trials (RCT)
   Learning to predict an unobserved target variable;
   - Spurious correlation; variables seem correlated but are not.

Question 3

Question 2
What is the fundamental problem of causal inference and discuss on a solution to this problem.

Answer 3

fundamental Challenge: learning to predict an unobserved target variable!
~ No comparison of prediction vs. observed outcome!

not causal inference (rather evaluation problem)

solution; randomized controlled trials (RCT)
aka ‘…all else equal…’: to isolate the (average) effect of the treatment from confounders

other scientific apporaches;
- ATE aka A/B testing (avg treatment effect)
- ITE aka lift modelling (indiv. treatm. effect)

Question 4

Question 3
What are individual treatment effects, and how are they different from average treatment effects?

Answer 4

ITL= effect of treatment on an individual unit
ATE= overall effect by averaging the individual treatment effects across all units in the population.

Question 5

Question 4
What is the difference between operations research, predictive analytics and prescriptive data analytics?

Answer 5

• Operations research (OR) – Prescriptive analytics
  Using mathematical modeling for developing simulation models to optimize decision variables
• Predictive analytics
  Using data and machine learning for developing predictive models
• Bridge? Prescriptive data analytics?
  Can data tell us what the optimal decision is? suggests actions to optimize results