How to Boost Beauty Brand Repeat Purchases with AI‑Driven Uplift Modeling

Scenario Introduction

With the rise of social‑media e‑commerce and beauty‑related KOLs/KOCs, online beauty shopping has surged, changing consumer buying behavior. While online channels drive sales, offline traffic and deep member operations remain critical for brand trust and loyalty.

Acquiring a new customer costs 5–10 times more than retaining an existing one, so improving repeat purchase rates through precise, technology‑enabled marketing is essential.

Solution Introduction

The proposed solution builds a high‑potential member prediction model combined with tiered segmentation for fine‑grained marketing, aiming to increase ROI and overall conversion.

Audience Selection Methods

Three common methods are used to select target audiences:

Member‑tag matching : Simple, broad targeting based on attributes such as membership level, product preferences, or purchase amount.

Behavior‑based rule selection : Targets users with strong purchase‑related behaviors (e.g., page views, cart adds, offline activity).

Model‑based tiered selection : Uses machine‑learning models to predict purchase probability and selects high‑probability members for marketing.

Model Definitions

Natural Repurchase Model : Predicts the probability of purchase without any intervention. It treats the problem as a binary classification where the target variable indicates whether a member will purchase within a given time window.

Purchase Power Model : Extends the natural repurchase model to predict a member’s purchase‑level (high, medium, low) after a purchase occurs, helping identify high‑value customers.

Marketing Response Model : Predicts the probability of purchase after a marketing intervention, focusing on the uplift caused by the intervention.

Marketing Gain (Uplift) Model : Estimates the difference between the post‑intervention purchase probability and the natural repurchase probability, identifying users who are sensitive to marketing actions.

Dataset Construction

All models share a similar data pipeline:

Define the prediction problem (member ID, time window, target variable).

Build a sample dataset using historical data (e.g., past three months) with member IDs and time windows.

Construct feature datasets by extracting attribute, statistical, time, and cross features while avoiding data leakage.

Merge sample and feature datasets into a wide table for model training and prediction.

Model Building

Binary classification models such as Logistic Regression, SVC, KNN, Random Forest, or LightGBM can be used for natural repurchase, marketing response, and uplift models. Multi‑class classifiers are applicable for the purchase‑power model.

Model Evaluation

Because positive samples are often scarce, metrics like ROC‑AUC, PR‑AUC, AUUC (Area Under Uplift Curve), and Qini Coefficient are preferred over accuracy or log‑loss. Evaluation should consider monotonicity of uplift scores and the separation between experimental and control groups.

Comparison and Recommendations

Modeling difficulty (from easiest to hardest): Natural Repurchase < Marketing Response < Purchase Power < Marketing Gain. Business implementation difficulty follows a similar order, with uplift models requiring randomized experiments and balanced control/treatment groups.

For most brands, combining natural repurchase and marketing response models yields quick wins, while purchase‑power and uplift models provide deeper insights when data quality and experimental infrastructure are available.

Conclusion

There is no universally superior model; selection should align with specific business pain points and data availability. Continuous experimentation and data‑driven refinement are key to improving repeat purchase rates and overall marketing ROI.