Disentangled self-attention neural network based on information sharing for click-through rate prediction

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Recommendation systems combat information overload with growing network resources.
• Click-through rate (CTR) prediction is vital for intelligent recommendation systems.
• Existing CTR models use parallel networks but neglect unary terms and cause input oversharing.

Purpose of the Study:

• To propose a novel disentangled self-attention neural network (DSAN) for CTR prediction.
• To address limitations in existing models regarding unary term interaction and feature input oversharing.
• To simulate complex feature interactions more effectively.

Main Methods:

• Employing an embedding layer to convert high-dimensional sparse features into low-dimensional dense matrices.
• Utilizing disentangled multi-head self-attention to learn feature relationships.
• Implementing a shared interaction layer to enhance information sharing within parallel networks.

Main Results:

• The proposed DSAN method demonstrates superior predictive accuracy compared to existing models.
• Experiments on two real-world datasets validate the effectiveness of DSAN.
• DSAN successfully models complex feature interactions, including unary terms.

Conclusions:

• DSAN offers a significant advancement in CTR prediction accuracy.
• The model effectively addresses information overload in recommendation systems.
• DSAN provides a more sophisticated approach to feature interaction modeling.