BaGFN: Broad Attentive Graph Fusion Network for High-Order Feature Interactions

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Effective modeling of feature interactions is vital for high-quality feature engineering, especially with multifiled sparse data.
• Current state-of-the-art methods often extract cross features implicitly, limiting their ability to capture complex interactions across different feature fields.

Purpose of the Study:

• To propose a new Broad Attentive Graph Fusion Network (BaGFN) for flexible and explicit modeling of high-order feature interactions.
• To enhance the comprehensive competence of learning sophisticated interactions among diverse feature fields.

Main Methods:

• Designed an attentive graph fusion module utilizing a novel bilinear-cross aggregation function and self-attention for high-order feature representation under graph structures.
• Developed a broad attentive cross module with a new broad attention mechanism to refine feature interactions at a bitwise level, dynamically learning importance weights.

Main Results:

• The proposed BaGFN effectively models high-order feature interactions in a flexible and explicit manner.
• Experimental results demonstrate the superior effectiveness of the BaGFN compared to existing methods.

Conclusions:

• The BaGFN offers a significant advancement in modeling complex feature interactions for multifiled sparse data.
• The model's ability to explicitly learn high-order feature relationships enhances feature engineering capabilities.