Causal machine learning for personalized discount targeting with technographic trace data
Causal Machine Learning
Gradient Boosted Trees
Singular Value Decomposition
Off-Policy Evaluation
Explainability
Conference Awards
Project Overview:
I built a profit-oriented targeting framework for online discounts that learns optimizes over treatment effect heterogneity directly from A/B-test data. My framework optimizes both the hyperparameter tuning and targeting algorithm training for economically relevant objective functions, as opposed to off-the-shelf statistical objectives. Using data from 148k website sessions from two campaigns, I show how the method boosts expected profits by 3-6 % versus best-practice baselines, entirely with technographic trace data (device, screen, browser, etc.).

Read the Research Paper
Manuscript with methodology, pseudocode for training algorithms, and empirical analysis
Key Contributions & Technical Execution
🎯 Research Leadership
Collaborated with experimentation platform to secure data, developed the research idea, conducted all analysis in Python, and authored research paper.
🧮 Algorithm Design
Combined SVD feature engineering with LightGBM gradient-boosted trees to estimate heterogeneous profit lift in a model that allows for flexible discount/cost structures.
🎲 Counterfactual Evaluation
Implemented doubly-robust off-policy evaluation and Monte-Carlo cross-validation to evaluate methodology.
🔍 Explainability
Developed grouped permutation importance technique to show device-level variables (screen size, OS, browser) drive most of the profit signal.
Key Research Findings
Economic vs. Statistical Targeting
Shows why profit-oriented criteria must balance baseline purchase rates and treatment effects—classic uplift rules leave money on the table. Project demonstrates the importance of optimizing both tuning and training algorithms for appropriate economic objectives.
3-6 % Profit Uplift
Optimal policy outperforms both uniform and uplift baselines across two firms. Provides clear quantification of the value of modern causal ML in real-world discount targeting.
Device-level characteristics matter most
While geographic and behavioral variables do contribute to the accuracy of the targeting algorithms, basic device-level information (screen size, operating system, browser) drive most of the incremental value in profit lift.
Recognition & Impact
🎙️ Conference Presentations
- Conference on Information Systems & Technology – Virtual
- Conference on Statistical Challenges in E-commerce Research – Virtual
- International Conference on Information Systems, Virtual
🏆 Best Short Paper in Track, 2nd Runner-up
International Conference on Information Systems, 2020
Based on early version of this work linked here (.pdf)