Large-Scale Notification Dispatch with Bundle Treatments and Multi-Outcome Uplift Optimization

Notification dispatch plays a critical role in large-scale user engagement, involving complex trade-offs across notification timing, presentation style, multiple outcomes, and constraints. In this paper, we formulate it as a constrained optimization over bundle treatments that jointly specify timing and presentation style, aiming to maximize incremental Daily Active Users (DAU) subject to platform-level budget and device vendor-specific quota constraints. The problem is challenging due to multi-dimensional, small-effect uplift estimation and large-scale constrained optimization. To address these challenges, we propose $\textbf{B}$undle $\textbf{U}$plift $\textbf{O}$ptimization with $\textbf{P}$runed $\textbf{L}$agrangian-based $\textbf{R}$elaxation (BUOPLR), a two-stage notification dispatch method that decouples uplift estimation from constrained decision-making. BUOPLR first learns bundle-level, multi-outcome small uplift through an architecture that captures cross-treatment and cross-outcome relationships, and then performs scalable assignment by restricting the decision space and applying Lagrangian relaxation to a small set of global constraints. Offline experiments show BUOPLR outperforms state-of-the-art methods, and online A/B tests increase DAU by 0.5\%. BUOPLR is now deployed on Kuaishou, a large-scale short video platform serving over 100 million users daily.

App platforms like short-video services send millions of push notifications daily, but deciding when and how to notify each user is far from trivial. The decision must simultaneously account for notification timing, presentation style, platform spending limits, and per-user quotas set by phone manufacturers—a resource allocation problem with billions of constraints that existing methods cannot solve efficiently. Moreover, the individual effect of each notification on user engagement is tiny and noisy, making accurate estimation extremely difficult. We developed BUOPLR, a two-stage approach that decouples effect estimation from decision-making. First, a custom model learns the small, incremental impact of each notification bundle—a combination of timing and style—across two outcomes: user activity and cost, capturing how richer notification styles build upon simpler ones and how engagement and cost relate. Second, a scalable solver ranks candidates by a score that balances benefit against cost competitiveness, aggressively prunes low-quality options, and reduces billions of constraints to just two global ones—making billion-user-scale optimization solvable within hours instead of days. BUOPLR is now deployed on Kuaishou, a platform serving over 100 million daily users, and has increased daily active users by 0.5% in online experiments while respecting all platform constraints.