In online advertising, predicting the likelihood of a user clicking on an ad is crucial for maximizing revenue. In this article, we explore the idea of context-aware ranking for click-through rate (CTR) prediction, which considers the relevance of the ad to the user’s context. We propose a novel approach called calibration-compatible listwise distillation (CLID), which combines the strengths of both ranking and classification models to improve CTR prediction performance.
Contextual Features
In online advertising, context is everything. The items that are displayed together with the candidate item on the same page are privileged ones for the ranking model when predicting the user’s probability of clicking the candidate item. These items and their features are called contextual features, which significantly affect the user’s click propensity. Therefore, properly utilizing private contextual features for the ranking model can improve CTR prediction performance.
CLID Approach
Our proposed CLID approach combines the strengths of both ranking and classification models. In the re-ranking stage, we use a classification model to predict the user’s probability of clicking the candidate item. Then, we use the contextual features to refine the predictions of the classification model. By doing so, we can better utilize the privately held contextual features for improving CTR prediction performance.
Benefits
The proposed CLID approach offers several benefits over traditional ranking models. Firstly, it can handle non-privileged features in the re-ranking stage without any issue of training-serving inconsistency. Secondly, it can significantly improve CTR prediction performance by utilizing privately held contextual features. Finally, it can provide a more accurate and robust ranking model for online advertising.
Conclusion
In conclusion, context-aware ranking is a crucial aspect of online advertising, and our proposed CLID approach offers a novel solution for improving CTR prediction performance. By combining the strengths of both ranking and classification models, we can better utilize privately held contextual features to refine predictions and provide a more accurate and robust ranking model. This approach has significant implications for online advertising platforms, as it can help improve revenue and user satisfaction by providing more relevant ads.
