LiftLab
causal-inference
pytorch
machine-learning
An uplift model that predicts who an ad will actually persuade, compared against a two-model baseline and random targeting.
Overview
Showing an ad to someone who would have bought anyway wastes money. Uplift modelling predicts the incremental effect of an ad on each person, so spend goes to the people it actually changes.
LiftLab trains a treatment-aware neural network on randomized advertising data. A dashboard compares three targeting strategies: the uplift model, a two-model baseline, and random targeting, each with confidence intervals.
Results
- Held-out AUUC rose from 0.100 to 0.118 (+18%) over the two-model baseline.
- Targeting the top 20% of customers ranked by predicted effect gave a 3.2 percentage-point incremental conversion lift over random targeting, measured on an untouched test set.
Making the numbers trustworthy
Uplift estimates are easy to fool yourself with, so the evaluation pipeline includes:
- Runs across five random seeds
- Bootstrap confidence intervals
- Treatment-balance checks
- Feature ablations to catch data leakage
Stack
Python · PyTorch · scikit-learn · Pandas · NumPy