Introduction
Modern e-commerce platforms face a persistent challenge: balancing short-term revenue from advertisements with long-term user satisfaction derived from organic product recommendations. Traditionally, these two systems operate in silos. A fixed number of ad slots are auctioned off, and the remaining space is filled by a recommendation engine. This "separate-then-integrate" approach often results in sub-optimal outcomes, as it fails to account for the externalities organic items impose on ad competition and vice versa.
To address this, researchers presented JINTER Net (Joint INTEgrated Regret Network) at AAAI-25. JINTER Net is a neural network architecture designed to jointly optimize the placement of both ads and organic items, treating auction design as an Automated Mechanism Design (AMD) problem.
Methodology: The JINTER Net Architecture
JINTER Net moves away from manual formula derivation, instead "learning" optimal allocation and payment rules that maximize a weighted sum of platform revenue and Gross Merchandise Volume (GMV), a proxy for user experience.
The architecture comprises two main components:
- Allocation Network: This network determines the probability of an item (ad or organic) occupying a specific slot. It employs a double-softmax mechanism: a row-wise softmax ensures slots are not over-allocated, while a column-wise softmax prevents item duplication.
- Payment Network: For ads, this network calculates payments that satisfy Individual Rationality (IR), ensuring advertisers are never charged more than their stated value. It accounts for the presence of organic items—treated as "passive bidders" with zero bids—by learning a scaling parameter for externalities.
The model is trained using an augmented Lagrangian method within the RegretNet framework. This ensures the system is approximately Dominant Strategy Incentive Compatible (DSIC), meaning advertisers achieve the best results by bidding their true valuations.
Experimental Evaluation
The authors evaluated JINTER Net using both synthetic data (Uniform, Normal, and Lognormal distributions) and the real-world Avito search log dataset, which includes over 112 million page views.
Key Results
- Performance Gains: JINTER Net consistently outperformed traditional baselines, including the Generalized Second Price (GSP) auction and the Integrated Auction System (IAS). On the Avito dataset, it achieved significant improvements in the combined Revenue + GMV metric.
- The Pareto Frontier: By adjusting a hyperparameter ($\alpha$), platforms can navigate the Pareto curve to find the ideal balance between immediate revenue and long-term user satisfaction.
- Economic Stability: The "regret" (the incentive for advertisers to misreport their values) remained below 0.001 across tests, confirming the system's practical truthfulness.
Conclusion and Future Work
JINTER Net demonstrates the power of neural networks in solving complex mechanism design problems where traditional theories struggle. By breaking the "fixed slot" constraint and integrating organic items into the auction framework, it offers a more flexible and globally optimal solution for e-commerce platforms.
Future research directions include relaxing the assumption of separable Click-Through Rates (CTR) to account for complex item-to-item interactions and improving the scalability of the AMD approach for real-time deployment in high-volume environments.
Reference: Ma, Y., Li, W., Wang, Y., et al. (2025). On Designing the Optimal Integrated Ad Auction in E-commerce Platforms. Proceedings of the 39th AAAI Conference on Artificial Intelligence (AAAI-25).
Summary updated