Machine Learning (algorithm)
Few-shot Learning for CTR/CVR Estimation in Online Advertising
Click-through rate (CTR) prediction and conversion rate (CVR) prediction are critical tasks in online advertising system. Plenty of businesses depend on CTR/CVR estimation, such as optimized cost-per-click (oCPC) bid strategy, pacing control and optimization of campaigns, smart ad creative selection, gross merchandise volume (GMV) optimization, etc. Thus the performance of CTR/CVR prediction has great impact on the revenue and user experience, and benefits the advertising system for its performance promotion.
Despite of the huge attention on CTR/CVR prediction from both academia and industry communities, there exists several task-specific problems that make CTR/CVR estimation challenging to handle. For instance,
- Lack of labeled samples in specific scenarios: E-commerce marketing campaigns and advertising materials are changing frequently. Although huge samples can be collected form online systems, there are only a few labeled data of specific marketing scenarios or new ads. And in the task of conversion rate prediction, there exist problems of delayed feedback and sparse positive samples;
- Cold-start problem: In Taobao online advertising system, there are a large number of new advertisements every day, especially for Alibaba's Double Eleven Sale (11/11). It is challenging to estimate CTR/CVR of these new ads;
- Costly hyperparameter tuning: It is costly to tune proper hyperparameters of CTR/CVR models for different similar scenarios.
Zero/ Few-shot learning refers to the technique of training model with zero or a small amount data, contrary to the normal practice where a large amount of data is used. There are a lot of works about few-shot learning in the field of computer vision, but not for CTR/CVR prediction. We hope to mitigate the aforementioned challenges by utilizing techniques like Zero/ Few-shot learning.
- Achieve better performance than state-of-the-art in CTR/CVR estimation of specific marketing scenarios or new ads
- Model/ Framework that can be applied to different tasks effectively
- Impactful research paper publication
Related Research Topics
- Zero/One/Few shot learning / Meta-learning for deep networks
- Deep learning for search engines, recommendations and online advertising
- Sample selection bias correction theory / Counterfactual evaluation and learning
- Multi-task models for CTR/CVR estimation
Suggested Collaboration Method
AIR (Alibaba Innovative Research), one-year collaboration project.