Introduction
In online advertising, Click-Through Rate (CTR) prediction is a critical task. However, it is plagued by a chronic class imbalance problem: users click on very few ads, resulting in datasets where non-clicks (negatives) vastly outnumber clicks (positives), often by ratios exceeding 5:1. Standard machine learning models struggle with this disparity. Down-sampling negatives discards information, while over-sampling positives with synthetic methods like SMOTE often fails to capture the complex, heterogeneous nature of tabular ad data.
The DAN Framework
To address these challenges, the authors propose the Disguise Adversarial Network (DAN). Unlike traditional Generative Adversarial Networks (GANs) that generate data from random noise—a process that often fails for structured tabular data—DAN transforms real negative samples into potential positives.
The framework consists of two main components:
- Disguise Neural Network ($T$): This network takes a real negative sample and applies "Ads Makeup"—subtle feature modifications—to create a "disguised" sample ($z$) that mimics a positive click. An $\ell_1$ distance penalty ensures these modifications remain realistic.
- Discriminator Neural Network ($D$): Serving as the actual CTR predictor, it is trained on real clicks, real non-clicks, and disguised samples.
The adversarial training employs a "mild objective" based on discriminative clustering. Instead of forcing hard labels on disguised data, the model minimizes conditional entropy, encouraging the discriminator to find a clear "info margin" between classes.
Experimental Setup
The model was evaluated on two large-scale industry datasets:
- Criteo (Display Ads): 46 million logs over 7 days.
- Avazu (Mobile Ads): 40 million logs over 10 days. A sliding window approach was used to simulate real-world temporal dynamics in ad delivery.
Performance and Robustness
DAN consistently outperforms a wide range of baselines, including standard Deep Neural Networks (DNN), Support Vector Machines (SVM), and specialized imbalance-handling techniques like SMOTE and ADASYN.
Key findings include:
- Superior AUC: On the Criteo dataset, DAN achieved an AUC of 0.75, significantly higher than the 0.69 achieved by a standard DNN.
- Data Efficiency: DAN is remarkably robust to data scarcity. When training data was reduced to just 10% of the original volume, DAN's AUC only dropped from 0.75 to 0.70, whereas other models experienced much steeper declines.
- Effective Augmentation: By using real negatives as "seeds," DAN effectively learns the manifold of ad data better than noise-based GANs like SGAN.
Conclusion and Limitations
DAN offers a powerful solution for CTR prediction under heavy class imbalance by leveraging "disguised" data to sharpen decision boundaries. However, the approach comes with increased training complexity, taking approximately 3x longer to train than a standard DNN. Future work may explore scaling this architecture to multi-class engagement scenarios and improving recommendation frequency.
Summary updated