How would you target promotions to grow consumers?
Company: DoorDash
Role: Machine Learning Engineer
Category: Machine Learning
Difficulty: medium
Interview Round: Onsite
## ML / Growth Scenario
You own a system that sends **promotion offers** (e.g., "$10 off", "free delivery") to consumers to increase growth.
### Prompt
Design an approach to decide:
- **Who** should receive a promo
- **What** promo to send (offer type/amount)
- **When** to send it
Constraints and concerns:
- Limited promo budget
- Risk of giving discounts to users who would have ordered anyway
- Need to measure **incremental lift** (not just conversion)
- Avoid harming long-term margin or creating promo addiction
Explain your modeling/experimentation strategy, data needs, evaluation metrics, and how you would deploy and monitor it.
Overview: This question evaluates the ability to design and operationalize ML-driven promotion targeting and measurement systems, assessing competencies in causal inference and uplift estimation, experimentation design, metric selection, budget-aware optimization, and monitoring for long-term impact.
Community answers
Answer by swemleds
Problem Definition
Frame this as a casual targeting and budget allocation problem.
goal: maximize incremental lift and contribution profit by promotion (not limited to order conversion)
Data
collect data including order behavior, prior promotion exposure, redemption, and contextual information
take randomized experiment with no-promotion control group to estimate treatment effect
Model
T-Learner with two LightGBM Models:
one for with-promo, the other for without promo; compute probability of purchase to get uplift difference
this can identify users whose ordering behavior are most likely to change with promo
to decide the amount of promo, we can estimate separately for different offer type e.g. free delivery, fiver dollors off, etc
for timing, we can integrate time of day, prior exposure, or ordering patterns
Cost Optimization
calculate expected incremental contribution profit by combining predicted uplift, order margin, and promo cost
can allocate limited budget on highest-cost promowhile enforcing eligibility and frequency caps
Evaluation
randomized A/B Test and focus on incremental orders, incremental contribution profit, and return on promo spend
long term retention and whether customers were dependent on promo could also be monitored
Deployment
deploy the system on a scheduled batch pipeline, maintain persistent control group
continuously monitor the impact and model performance