AIE: Auction Information Enhanced Framework for CTR Prediction in Online Advertising

Yang Yang,Bo Chen,Chenxu Zhu,Menghui Zhu,Xinyi Dai,Huifeng Guo,Muyu Zhang,Zhenhua Dong,Ruiming Tang

DOI: https://doi.org/10.1145/3640457.3688136

2024-08-15

Abstract:Click-Through Rate (CTR) prediction is a fundamental technique for online advertising recommendation and the complex online competitive auction process also brings many difficulties to CTR optimization. Recent studies have shown that introducing posterior auction information contributes to the performance of CTR prediction. However, existing work doesn't fully capitalize on the benefits of auction information and overlooks the data bias brought by the auction, leading to biased and suboptimal results. To address these limitations, we propose Auction Information Enhanced Framework (AIE) for CTR prediction in online advertising, which delves into the problem of insufficient utilization of auction signals and first reveals the auction bias. Specifically, AIE introduces two pluggable modules, namely Adaptive Market-price Auxiliary Module (AM2) and Bid Calibration Module (BCM), which work collaboratively to excavate the posterior auction signals better and enhance the performance of CTR prediction. Furthermore, the two proposed modules are lightweight, model-agnostic, and friendly to inference latency. Extensive experiments are conducted on a public dataset and an industrial dataset to demonstrate the effectiveness and compatibility of AIE. Besides, a one-month online A/B test in a large-scale advertising platform shows that AIE improves the base model by 5.76% and 2.44% in terms of eCPM and CTR, respectively.

Information Retrieval

What problem does this paper attempt to address?

### Problems the Paper Attempts to Solve

The paper "AIE: Auction Information Enhanced Framework for CTR Prediction in Online Advertising" attempts to address two major issues faced in click-through rate (CTR) prediction in online advertising recommendation systems:

1. **Insufficient Utilization of Auction Signals**:
- The online advertising auction environment is highly complex, and auction information (such as market price, CPC bids, etc.) significantly impacts CTR prediction. However, existing CTR prediction models often overlook these posterior auction signals, leading to suboptimal prediction performance.
- Although some existing studies have tried to incorporate posterior auction information, they have failed to fully exploit the potential value of this information, especially regarding the variability in market price distribution across different scenarios.

2. **Auction Bias**:
- In ad recommendations, ads with higher bids are more likely to win auctions, even if their actual click relevance is low. This results in these ads being overexposed to uninterested users, thereby reducing the CTR.
- This phenomenon creates a bias in the CTR distribution of the training data compared to the actual target distribution, known as "auction bias." This bias affects the accuracy of CTR prediction models.

To address these issues, the authors propose a new framework—Auction Information Enhanced Framework (AIE), which improves CTR prediction by introducing two pluggable modules:

- **Adaptive Market-price Auxiliary Module (AM2)**: By constructing auxiliary tasks that utilize market price information, this module dynamically captures the changes in market price distribution across different scenarios, enhancing the performance of the CTR prediction model.
- **Bid Calibration Module (BCM)**: By effectively utilizing bid information, this module corrects the auction bias in the training data, making the model's predictions more accurate.

Through these methods, AIE aims to fully leverage auction information and reduce auction bias, thereby improving the accuracy of CTR prediction and the revenue of the advertising platform.