TL;DR
This paper introduces MIAA, a novel deep learning framework that integrates ad auction and allocation for e-commerce feeds. By jointly optimizing these stages and accounting for externalities, MIAA significantly boosts platform revenue and Gross Merchandise Volume (GMV) while ensuring fairness for advertisers.
What Problem Does This Paper Address?
In e-commerce platforms, user-facing feeds display a mix of organic product listings and advertisements. The way these ads are auctioned and then placed within the feed directly impacts the platform's revenue and the overall sales volume (GMV). Traditionally, ad auction and allocation are treated as separate steps. This separation leads to two major issues:
- Ignoring Externalities: Ad auctions often don't consider how an ad's position and surrounding items (context) affect its click-through rate (CTR). This means the auction might not accurately reflect an ad's true potential value.
- Incentive Incompatibility (IC): The allocation stage, which uses the auction's outcome to place ads, can disincentivize advertisers from bidding higher. If increasing a bid doesn't guarantee a better position or higher utility, advertisers have less reason to bid their true value, leading to suboptimal outcomes for both advertisers and the platform.
This paper aims to bridge this gap by creating a unified mechanism that addresses these issues, ensuring fairness (Individual Rationality - IR, and Incentive Compatibility - IC) while maximizing platform revenue and GMV, even when considering complex interactions between ads and organic items.
Methods
The proposed solution, MIAA (Deep Automated Mechanism Design for Integrating Ad Auction and Allocation in Feed), is a deep learning framework with three core modules:
- Externality-aware Prediction Module (EPM): This module uses a list-wise prediction model to capture "global externalities." Instead of predicting CTR for each item in isolation, it considers the entire list of ads and organic items to predict the CTR and GMV for each position. This is crucial for understanding how an ad's placement and its neighbors influence user behavior.
- Automated Auction Module (AAM): Inspired by Automated Mechanism Design (AMD) and VCG-like mechanisms, AAM uses deep neural networks (µ-network and λ-network) to model auction parameters. This allows for a more expressive and flexible auction design that can guarantee IC and IR properties, even with externalities. It essentially learns how to set bids and payments optimally.
- Differentiable Sorting Module (DSM): To enable end-to-end training, DSM uses a differentiable relaxation of the sorting operator (via softmax) to select the winning allocation. This allows gradients to flow back through the entire system, enabling the optimization of all modules simultaneously.
MIAA jointly determines the ad's ranking, payment, and display position, creating a holistic optimization process.
Datasets
The paper uses two datasets for evaluation:
- Avito: A public dataset from a Kaggle competition for context ad clicks. It's used for simulating user click behavior and evaluating the list-wise prediction model.
- Meituan: An industrial dataset collected from the Meituan retail delivery platform. This dataset is used for both offline experiments and online A/B tests, providing real-world performance data.
Key Insights and Ideas
- The Power of Joint Optimization: Separating auction and allocation is suboptimal. Integrating them allows for a more accurate reflection of an ad's true value and a better balance between revenue and user experience.
- Externalities Matter: The position and context of an ad significantly influence its performance. List-wise prediction models are essential for capturing these complex interactions, outperforming traditional point-wise models.
- Deep Learning for Mechanism Design: Leveraging deep neural networks within an AMD framework allows for more sophisticated and adaptive auction and allocation strategies that can handle complex externalities and guarantee economic properties like IC and IR.
- End-to-End Learning for Real-World Impact: The differentiable sorting module is key to enabling end-to-end training, allowing the entire system to learn and adapt based on predicted outcomes, leading to significant improvements in practical applications.
Key Results
- Improved Prediction Accuracy: The list-wise EPM module significantly improves AUC and PCOC (a measure of prediction accuracy) for CTR prediction compared to point-wise models on both public and industrial datasets. For example, on the Meituan dataset, AUC improved from 0.6485 to 0.7077 for all positions.
- Significant Revenue and GMV Gains: Offline experiments showed MIAA outperforming baseline mechanisms like GSP+Fixed Position, GSP+Cross DQN, Score-Weighted VCG, and IAS. On the Avito dataset, MIAA achieved a 5.35% lift in Revenue+αGMV.
- Online A/B Test Success: Online A/B tests on the Meituan platform demonstrated that MIAA achieved substantial improvements across all performance metrics (revenue, GMV, RPM, GPM) compared to GSP+Fixed Position and GSP+Cross DQN. For instance, MIAA showed an 8.17% improvement in Revenue+αGMV compared to GSP+Fixed Position.
Limitations and Future Work
- Limitations: The paper acknowledges that the offline training process relies on predicted data and doesn't have direct access to real user behavior feedback, which can lead to inconsistencies between offline and online performance. The current work focuses on single-slot ad allocation; integrating multi-slot ads is a more complex challenge.
- Future Work: The authors suggest investigating efficient mechanisms for integrating multi-slot ads auction and allocation. Further research could also explore more robust methods for handling the offline-to-online performance gap.
Summary updated