Uplift modeling papers
Tight starter set — the causal turn (Q3): who converts because of an intervention (uplift/CATE), not who'd convert anyway (propensity). Feeds Attribution / uplift and Causal inference primer. OA legend: ✅ open · 🟡 free copy · 🔒 paywalled. Read with How I read a paper.
🎯 Start here:
gutierrez2017uplift— the only survey here; maps every method family so the rest land in context. Path: gutierrez → athey → wager → kunzel → rzepakowski → anderl
⬜ P1 · Causal Inference and Uplift Modelling: A Review of the Literature
Gutierrez & Gérardy (2017) · PMLR v67:1–13 · survey · gutierrez2017uplift · ✅ yes
🔗 https://proceedings.mlr.press/v67/gutierrez17a.html
Why: Maps the three uplift families (two-model, class-transformation, direct) and frames predict-vs-cause explicitly — why propensity ≠ uplift (Predict vs cause checklist).
⬜ P2 · Recursive Partitioning for Heterogeneous Causal Effects
Athey & Imbens (2016) · PNAS 113(27):7353–7360 · method · athey2016recursive · ✅ yes
🔗 https://www.pnas.org/doi/abs/10.1073/pnas.1510489113
Why: Introduces "honest" causal trees — split the partition-building and estimation samples for valid CATE inference. The key idea that separates causal partitioning from ordinary trees.
⬜ P3 · Estimation and Inference of Heterogeneous Treatment Effects using Random Forests (Causal Forests)
Wager & Athey (2018) · JASA 113(523):1228–1242 · method · wager2018causal · ✅ yes (arXiv)
🔗 https://arxiv.org/abs/1510.04342
Why: Scales honest trees to an ensemble with consistency + confidence intervals on CATEs (the grf package). Key idea: weight neighbours by treatment-assignment similarity.
⬜ P4 · Metalearners for Estimating Heterogeneous Treatment Effects using Machine Learning
Künzel, Sekhon, Bickel & Yu (2019) · PNAS 116(10):4156–4165 · method · kunzel2019metalearners · ✅ yes
🔗 https://pmc.ncbi.nlm.nih.gov/articles/PMC6410831/
Why: Defines S/T/X-learners; the X-learner shines under imbalanced treatment/control (typical in marketing). Extract its which-strategy-when decision logic as a checklist.
⬜ P5 · Decision Trees for Uplift Modeling with Single and Multiple Treatments
Rzepakowski & Jaroszewicz (2012) · Knowledge and Information Systems 32(2):303–327 · method · rzepakowski2012uplift · 🟡 free PDF
🔗 https://link.springer.com/content/pdf/10.1007/s10115-011-0434-0.pdf
Why: The canonical "model uplift directly" paper — divergence-based splits (KL, Euclidean, χ²) on treatment-vs-control distributions instead of two-model subtraction.
⬜ P6 · Mapping the Customer Journey: Graph-Based Online Attribution Modeling
Anderl, Becker, von Wangenheim & Schumann (2016) · Int. J. Research in Marketing 33(3):457–474 · method · anderl2016mapping · 🟡 SSRN
🔗 https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2685167
Why: Markov-chain multi-touch attribution via removal effects — the formalism that links attribution to causal reasoning, benchmarked against last-click/linear heuristics (Attribution models).