TL;DR
This paper introduces LBCF, a novel algorithm for optimizing marketing incentive allocation under budget constraints. By combining a unified causal forest for estimating treatment effects and an efficient dual gradient bisection method for budget optimization, LBCF significantly improves campaign engagement on a large-scale video platform, outperforming existing methods.
What Problem Does This Paper Address?
Online platforms frequently use incentives like discounts or bonuses to boost user engagement and revenue. However, these incentives come at a cost, and their effectiveness can vary greatly among users (heterogeneous treatment effects). Simply offering the largest incentive to everyone or using a greedy approach can lead to wasted budget and suboptimal returns. The core challenge is to determine the optimal amount and type of incentive to offer to each individual user, given a fixed overall budget, to maximize the platform's return on investment. This is framed as the "budget-constrained treatment selection" (BTS) problem. Existing methods struggle with the scale of modern platforms and often fail to accurately compare treatment effects across different incentive levels.
Methods
The paper proposes a two-pronged approach:
- Large-Scale Budget-Constrained Causal Forest (LBCF):
- Unified Discriminative Causal Forest (UDCF): Unlike previous methods that train separate models for each incentive level (leading to incomparable treatment effect estimates), UDCF builds a single causal forest. This unified model allows for direct comparison of treatment effects across multiple incentive levels within the same feature space. It uses novel splitting criteria (Inter and Intra splits) to encourage both inter-node and intra-node heterogeneity, crucial for effective treatment selection.
- Dual Gradient Bisection (DGB): To address the computational challenges of optimizing incentive allocation across millions of users and potentially many incentive options, the authors propose DGB. This is an efficient, parallelizable algorithm that leverages the convexity of the problem's dual formulation to find the optimal budget allocation. It's designed to work well with distributed computing frameworks.
- Percentage Mean Gain (PMG) Policy Evaluation: A novel offline evaluation metric is introduced to better assess the performance of BTS algorithms using data from Randomized Control Trials (RCTs). PMG aims to provide a more holistic measure of a policy's effectiveness compared to existing metrics that may not accurately reflect the true outcome or budget usage.
Datasets
- Synthetic Dataset: Used for simulation analysis, generated using methods from prior work ([26]) with varying noise levels.
- Real-world RCT Dataset: Collected from a large-scale video streaming platform. This dataset contains user engagement data (campaign engagement duration) across seven different incentive groups (multi-treatment) and a control group. The paper notes that while the full dataset is not publicly released due to privacy concerns, a subset is available for code testing.
Key Insights and Ideas
- Unified Causal Forests are Crucial for Multi-Treatment Comparison: The paper highlights a critical flaw in prior multi-treatment approaches: training separate models leads to incomparable treatment effect estimates. LBCF's unified model directly addresses this, enabling more accurate comparisons and better decision-making.
- Greedy Treatment Selection is Sub-optimal: The authors demonstrate with an example (Table 1) that simply picking the treatment with the highest individual ROI or uplift for each user is not optimal when a global budget constraint is in place. This underscores the need for a more sophisticated optimization strategy.
- Efficient Optimization for Large-Scale Systems: Traditional optimization methods for knapsack-like problems are not designed for modern distributed systems. DGB offers a scalable and efficient solution, making budget-constrained optimization feasible for platforms with millions of users.
- A Better Way to Evaluate BTS Policies: The PMG metric is a significant contribution, offering a more robust way to compare different incentive allocation strategies offline, which is a persistent challenge in this domain.
Key Results
- Simulation Analysis: LBCF consistently outperformed baseline methods (CTS, ED, Chi, CT.ST, CF.DT) across various noise levels, demonstrating its robustness.
- Offline Test: LBCF achieved the largest Percentage Mean Gain (PMG) for user campaign engagement duration across different budget configurations on the real-world dataset.
- Online AB Test: In a real-world deployment on a large-scale video platform, LBCF achieved a significant percentage gain in campaign engagement duration, consistently outperforming baseline methods like CT.ST and CF.DT. At the end of a 15-day test, LBCF showed a 7.79% gain, compared to 6.87% for CT.ST and 5.31% for CF.DT.
- Scalability: The LBCF algorithm is currently serving hundreds of millions of users on the platform, demonstrating its practical scalability.
Limitations and Future Work
- SUTVA Assumption: The authors acknowledge that the SUTVA (Stable Unit Treatment Value Assumption) assumption of independence between users might not strictly hold, as users might share incentive information. They mention that recent work exists to address such dependencies.
- Adding More Constraints: The authors suggest future work could involve incorporating additional constraints into the optimization task, such as user experience, and developing efficient solutions for these more complex scenarios.
Summary updated