uplift-modeling

Uplift models: S-learner baseline and causal forest

Trained on Criteo's uplift benchmark, 13,979,592 rows from a real randomized ad-exposure incrementality test. Full writeup and every gate: github.com/Bauxitiego/uplift-modeling.

Only two of the four methods evaluated are published here. T-learner and X-learner lost to the naive S-learner baseline on both targets tested - publishing them as ready-to-use artifacts would imply they're worth deploying, which the results say they aren't. S-learner (the baseline nothing but the causal forest beat) and CausalForest (the one method that did) are the two ends of the actual finding.

Results these models produced

3M training rows, 1M held out, same split for both methods. Causal forest beats the baseline on both targets - modestly (+1.8% visit, +3.8% conversion), but real and consistent in direction - at a real cost: 500-700s to train against 4s for the S-learner.

Loading a model

S-learner: plain XGBoost, saved in XGBoost's own native format. No pickle, nothing to disclose about trust.

from xgboost import XGBRegressor
model = XGBRegressor()
model.load_model("s_learner_visit/model.json")
# tau_hat(x) = model.predict(x with treatment=1) - model.predict(x with treatment=0)

CausalForest: joblib, not a safer format - checked directly, econml's CausalForest has Cython-based tree internals that skops cannot reconstruct (its Tree.__cinit__ needs constructor arguments skops has no way to supply). This is the one model in this repo that needs a trusted-source caveat on load. Only load causal_forest_*.joblib files from a source you trust, same as any pickle-based artifact.

import joblib
cf = joblib.load("causal_forest_visit/causal_forest.joblib")
tau_hat = cf.predict(X).flatten()

Each directory also has metadata.json with the exact train/test split size, seed, and feature list the model was fit with.

Honest scope

Meta-learners (Kunzel et al. 2019) and causal forests (Athey & Wager 2019) are established methods, not novel here. What's real: the meta-learners were implemented directly rather than called from a pre-built uplift library, validated against synthetic ground truth before being trusted on real data (where, by the fundamental problem of causal inference, no individual ground truth is ever available), and the evaluation metric itself was verified correct with dedicated tests before being relied on as the only judge.

License

Apache 2.0 for code and these model weights. The underlying Criteo data is not redistributed here - it's already published at its source, linked above, under its own terms.