Debiasing the Conversion Rate Prediction Model in the Presence of Delayed Implicit Feedback.

Area of Science:

• Machine Learning
• Recommender Systems
• Online Advertising

Background:

• Recommender systems (RS) are crucial for online advertisements, where conversion rate (CVR) prediction aids in evaluating ad impact and user profiling.
• Implicit feedback data is more prevalent than explicit feedback, but its direct use in RS can cause suboptimal performance due to inherent selection bias.
• Existing methods like reweighting address selection bias but often overlook delayed feedback, a common issue with limited observation times.

Purpose of the Study:

• To develop a novel approach for debiasing CVR prediction models in recommender systems.
• To effectively handle selection bias and delayed implicit feedback in real-world scenarios.
• To improve the accuracy and reliability of CVR prediction for online advertisements.

Main Methods:

• A novel likelihood approach combining a parametric model for delayed feedback with a reweighting method to mitigate selection bias.
• Minimizing a likelihood-based loss function utilizing multi-task learning.
• Evaluation on real-world datasets (Coat and Yahoo) to validate the proposed methods.

Main Results:

• The proposed methods demonstrated significant improvements in Area Under the Curve (AUC).
• Achieved a 5.7% AUC improvement on the Coat dataset and a 3.7% AUC improvement on the Yahoo dataset compared to baseline models.
• Successfully debiased the CVR prediction model, even with the presence of delayed implicit feedback.

Conclusions:

• The novel likelihood approach effectively addresses selection bias and delayed feedback in CVR prediction for recommender systems.
• The proposed methods offer a robust solution for improving the performance of online advertising platforms.
• This research contributes to more accurate and reliable user behavior modeling in recommender systems.