This project explores Uplift Modeling to measure the impact of two marketing offers:
- Discount (Treatment: Discount, Control: No Offer)
- Buy One Get One (BOGO) (Treatment: BOGO, Control: No Offer)
The goal is to identify persuadable customers, optimize targeting, and maximize conversion rates while minimizing marketing waste.
| Metric | Discount (Case A) | BOGO (Case B) |
|---|---|---|
| Total Conversions (Treated) | 1,877 | 1,450 |
| Total Conversions (Control) | 1,144 | 1,135 |
| Net Uplift | +733 | +315 |
| Negative Uplift Impact | -0.198 | -0.217 |
- Discount had a higher uplift than BOGO, leading to more conversions.
- BOGO showed positive uplift, but it was less effective than Discount.
- Negative uplift was observed in both cases, meaning some customers reacted negatively to the offer.
| Model | AUUC (Discount) | AUUC (BOGO) |
|---|---|---|
| S-Learner | 0.523 | 0.502 |
| Uplift Random Forest | 0.477 | 0.509 |
- S-Learner performed better for Discount, making it the best choice for uplift modeling in this case.
- Uplift Random Forest performed slightly better for BOGO, but its performance was inconsistent.
- Uplift Modeling Approaches:
- S-Learner (Meta-Learner using LightGBM)
- Uplift Random Forest
- Evaluation Metrics:
- AUUC (Area Under Uplift Curve)
- Gain Charts for Treatment Effectiveness
- Feature Engineering & Segmentation:
- Customer behavior analysis (Recency, Purchase History, Referral Impact)
- One-hot encoding for categorical variables
✔ Prioritize Discount over BOGO for promotions, as it led to 733 more conversions compared to 315 from BOGO.
✔ Use S-Learner for Discount campaigns, as it outperformed Uplift Random Forest (AUUC: 0.523 vs. 0.477).
✔ Improve segmentation to avoid targeting customers with negative uplift scores (-0.198 for Discount, -0.217 for BOGO).