Meta-Wrapper: Differentiable Wrapping Operator for User Interest Selection in CTR Prediction.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Recommender Systems

Background:

• Click-through rate (CTR) prediction is crucial for recommender systems.
• Deep learning models with attention mechanisms have improved CTR prediction by modeling user interests from historical behavior.
• Existing methods typically train attention modules jointly with base predictors using gradient descent.

Purpose of the Study:

• To propose a novel approach for user interest modeling, framed as a user interest selection problem.
• To introduce Meta-Wrapper, a method based on the wrapper feature selection framework.
• To enhance the performance and robustness of CTR prediction models.

Main Methods:

• Framing user interest modeling as a feature selection problem (user interest selection).
• Developing Meta-Wrapper using a differentiable module as a wrapping operator.
• Recasting the learning problem as a continuous bilevel optimization solved by a meta-learning algorithm.

Main Results:

• Theoretical convergence proof for the meta-learning algorithm.
• Demonstrated efficiency improvements in wrapper-based feature selection.
• Showcased enhanced resistance to overfitting compared to existing methods.
• Significant performance boosts in CTR prediction across three public datasets.

Conclusions:

• Meta-Wrapper offers a superior and more robust approach to user interest modeling for CTR prediction.
• The method effectively addresses feature selection challenges in recommender systems.
• Meta-Wrapper enhances the overall performance of CTR prediction models, outperforming existing techniques.