Uplift Modeling for Intelligent Marketing: Concepts, Methods, Evaluation, and Business Applications

1 Background

With the rapid development of the Internet and AI, intelligent marketing has become pervasive across industries. On the Beike platform, multiple consumer‑touch channels (DMP tagging, SMS/IM, coupons, DSP ads) are used, but a large portion of users convert naturally without any marketing spend. The key challenge is to measure and predict the incremental lift from marketing interventions and avoid wasting budget on users who would convert anyway.

The audience is divided into four quadrants: Persuadables (convert only with coupons), Sure thing (convert regardless), Lost cause (never convert), and Sleeping dogs (convert only without coupons). The goal is to target Persuadables while avoiding Sleeping dogs.

2 Value of Uplift Models in Intelligent Marketing

Traditional response models predict the probability of conversion after seeing a coupon but cannot distinguish whether the conversion is caused by the coupon or would have happened anyway. This is a causal inference problem. Uplift models predict the incremental effect of a treatment (e.g., an ad or coupon) on each individual:

P(Y=1|X) is the response model; P(Y=1|G,X) is the uplift model.

Illustrative example: two user groups have different raw conversion rates (0.8% vs 2.0%). When accounting for the baseline (no‑ad) conversion, the lower‑baseline group actually yields a higher uplift, showing why response models can be misleading.

3 Representation of the Uplift Model

For N users, let Y_i(1) be the outcome under treatment and Y_i(0) under control. The individual treatment effect (ITE) is τ_i = Y_i(1) – Y_i(0). The training objective is to maximize the expected uplift (CATE): τ(X_i) = E[Y_i(1) – Y_i(0) | X_i]. Because we never observe both outcomes for the same user, the observable formulation becomes:

Assuming conditional independence (CIA) between treatment assignment W_i and potential outcomes given features X_i, the CATE can be estimated from observed data.

4 Modeling Approaches

Four main families are discussed:

Two‑Model (Difference‑Response) Approach : Train separate models for treatment (G^T) and control (G^C) and subtract their predicted scores. Simple but error accumulates.

One‑Model (Difference‑Response) Approach : Append a binary treatment indicator T to the feature vector and train a single model, then predict twice (T=1 and T=0) and take the difference.

Label‑Transformation Method : Convert the uplift problem into a binary classification task by redefining the label Z_i (1 for treated‑converted or control‑non‑converted, 0 otherwise) and train any classifier on (X_i, Z_i).

Direct Uplift Modeling : Modify existing learners (LR, k‑NN, SVM, tree‑based models) to optimize uplift directly, e.g., by changing the split criterion to an uplift‑specific gain.

Each method’s pros and cons are summarized in tables and flow‑charts (see images).

5 Evaluation of Uplift Models

Standard classification metrics (AUC, precision, recall) are not applicable because true uplift is unobservable. Instead, the Qini curve (cumulative uplift) and AUUC (Area Under Uplift Curve) are used. The Qini curve is built by sorting users by predicted uplift, taking top‑ϕ fractions, and computing the difference in conversion rates between treatment and control groups. The area between the Qini curve and a random baseline quantifies model performance.

When treatment and control group sizes differ, a cumulative gain curve with a normalization factor (n_t(ϕ)+n_c(ϕ)) mitigates distortion.

6 Business Application on Beike New‑House Channel

The uplift model is used to identify coupon‑sensitive users when they land on the new‑house homepage, aiming to increase viewing and transaction rates while keeping ROI positive. A recent “618惠住季” campaign provides the experimental (coupon‑shown) and control (no coupon) groups. Positive samples are defined as users who, within 7 days after coupon exposure, performed any of: entrust, viewing, or purchase.

Sample distribution tables are shown, and the control group is down‑sampled to balance sizes.

Feature engineering includes behavior counts, city one‑hot, preference distribution, and LSTM‑encoded sequential features.

Models are trained with XGBoost as the base learner across the four approaches. The best performance is achieved by the One‑Model approach with treatment indicator features (no label transformation), where uplift is obtained by predicting twice (T=1, T=0) and subtracting.

Qini curves for each method illustrate that the two‑model approach performs worst, while the one‑model with propensity score adjustment also shows strong results.

7 Conclusion

The article summarizes the fundamentals, representations, modeling techniques, evaluation methods, and practical deployment of uplift modeling, emphasizing its high data‑quality requirements and potential extensions such as multi‑task learning and reinforcement learning for more complex marketing scenarios.