Push Notification Volume Optimization Using Uplift Modeling at Tencent Mobile QQ Browser

Introduction: The article presents Tencent's practice of optimizing QQ Browser push notification volume using an uplift model.

Push system characteristics: Push differs from recommendation systems as it is passive, may cause interference, and requires careful timing, quota, content relevance, and resource optimization.

Problem decomposition: The push workflow is broken into timing, cumulative impact, content‑interest matching, and resource optimization, focusing on cumulative impact where multiple pushes are treated as a single intervention.

Causal analysis: Traditional uplift models face challenges such as the lack of individual‑level counterfactual labels, high‑dimensional features drowning the treatment effect, and reliance on outcome calibration.

Improved modeling: Tencent introduced a modified S‑learner with pairwise ranking and metric learning to directly learn uplift, incorporating treatment features and multi‑task learning, resulting in reduced bias and smoother elasticity curves.

Results: The enhanced uplift model raised daily active users (DAU) by 2.8% compared with 1.38% from the baseline, and the work was submitted as a paper titled “Push Notification Volume Optimization Based on Uplift Model at Tencent Mobile QQ Browser”.

Q&A highlights: Discussed uplift model features, advantages over traditional methods, handling sparse treatment data, loss functions, and evaluation of ROI.