Abstract
In the mobile Internet era, mobile payment services have become a cornerstone of inclusive finance, offering convenience and security. Marketing strategies often encourage activity by allocating incentives like coupons to customers or commissions to merchants. However, two significant issues arise: first, the mutual influence between merchants and customers (bipartite influence) makes independent optimization sub-optimal. Second, coupon redemptions are only partially observed at specific incentive values, leading to a data censorship issue. This paper proposes a novel joint incentive optimization framework using Graph Neural Networks (GNNs) to model bipartite influences and hazard regression with a piecewise nonlinear function to capture response changes across incentive values. A linear programming method then allocates optimal incentives in real-time. Extensive results demonstrate the framework's effectiveness in real-world scenarios.
The Bipartite Incentive Problem
In mobile payment ecosystems like Alipay, marketing campaigns typically involve a two-sided incentive structure. As illustrated in Figure 1, a customer receives a coupon for scanning a QR code, while the merchant receives a commission for facilitating the transaction.
Traditionally, platforms optimize these incentives independently. This approach fails to account for two critical hurdles:
- Bipartite Influence: Merchants and customers are interdependent. High commissions motivate merchants to push more customers to scan codes, while high customer volume encourages merchants to stay engaged.
- Data Censorship: User price sensitivity is rarely fully observable. If a user redeems a $1.00 coupon, we do not know if they would have redeemed a $0.50 one (left-censored), or if they would have used a $1.50 one had they failed at $1.00 (right-censored).
Methodology: A Joint Optimization Framework
The researchers propose a three-part framework to address these challenges, integrating representation learning with survival analysis techniques.
Joint Representation Learning
The system builds a massive bipartite graph where nodes represent customers and merchants, and edges represent historical transactions. A GNN aggregates features from neighbors, allowing the model to capture how a merchant’s behavior influences their local cluster of customers and vice versa.
Response Modeling via Hazard Regression
Borrowing from medical survival analysis, the model treats the incentive value as the "time" variable. Hazard regression calculates the probability of a redemption occurring by a certain incentive value. This naturally enforces monotonicity—the principle that a higher incentive should never result in a lower probability of redemption.
Real-Time Optimization
Budget allocation is formulated as a linear programming problem. By using the Lagrangian multiplier method, the system derives a "shadow price" for the budget, allowing for millisecond-level optimal decisions across millions of requests.
Experimental Results
The framework was tested using a massive real-world dataset from Alipay, featuring over 112 million vertices and 133 million edges.
Key Insights
- Psychological Thresholds: The model detected significant "change points" in user behavior (e.g., at the 80-cent and 100-cent marks). Users are not linear calculators; they have specific thresholds where an incentive suddenly feels "worth it."
- Synergy Effect: Modeling merchants and customers together outperformed separate modeling. A better representation of a customer helps predict merchant performance because their fates are linked within the bipartite graph.
Online Performance
In live production A/B tests at Alipay, the framework achieved:
- Transaction Volume: A +2.54% relative improvement in total payments.
- Cost Efficiency: A -4.04% reduction in total marketing costs.
Conclusion and Future Work
While the framework effectively addresses bipartite influence and data censorship, it faces challenges with "cold start" scenarios for new users or merchants lacking transaction history. Future work will explore more complex graph structures, such as heterogeneous graphs, and apply this joint optimization to other two-sided markets beyond mobile payments.
Summary updated