Click-through Rate Prediction and Uncertainty Quantification Based on Bayesian Deep Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Click-through rate (CTR) prediction is crucial for recommendation systems and advertising traffic analysis.
• Existing deep learning models for CTR prediction often lack robust uncertainty quantification.
• Accurate modeling of uncertainty is essential for reliable real-world machine learning applications.

Purpose of the Study:

• To develop a CTR prediction framework that combines feature selection and interaction.
• To propose a Bayesian deep learning model for quantifying prediction uncertainty.
• To achieve accurate and reliable CTR predictions by addressing the limitations of deterministic models.

Main Methods:

• A novel CTR prediction framework integrating feature selection and interaction was designed.
• A Bayesian deep learning model utilizing Monte Carlo dropout was employed on a squeeze network and DNN parallel framework.
• Epistemic and aleatoric uncertainties were defined and quantified using information entropy and mutual information.

Main Results:

• The proposed Bayesian deep learning model demonstrated superior prediction performance compared to existing models.
• The framework successfully quantified both epistemic and aleatoric uncertainties in CTR prediction.
• Integrated prediction results were obtained by combining outputs from the parallel prediction model.

Conclusions:

• The developed Bayesian deep learning framework offers enhanced CTR prediction accuracy and reliability.
• The model's ability to quantify uncertainty provides valuable insights for practical applications.
• This approach advances the field of uncertainty modeling in machine learning for CTR prediction.