GAIN: A Gated Adaptive Feature Interaction Network for Click-Through Rate Prediction.

Area of Science:

• Machine Learning
• Data Mining

Background:

• Click-Through Rate (CTR) prediction is crucial for online advertising revenue.
• Existing models often struggle with the complexity of learning feature interactions, leading to suboptimal performance.

Purpose of the Study:

• To propose a novel, lightweight, and effective model for CTR prediction.
• To address the limitations of existing methods in handling feature interactions.

Main Methods:

• Introduced the Gated Adaptive feature Interaction Network (GAIN), a model featuring a novel cross module.
• The cross module utilizes gated units to learn arbitrary-order feature interactions, selectively dropping meaningless ones.
• Combined the cross module with a deep module for comprehensive feature interaction learning.

Main Results:

• GAIN achieved comparable or superior performance against state-of-the-art models on public datasets.
• Transferring GAIN's learned feature interactions significantly improved the performance of other models like Logistic Regression and Factorization Machines.

Conclusions:

• GAIN provides an effective and efficient approach to learning feature interactions for CTR prediction.
• The model's ability to learn and transfer feature interactions offers broad applicability in improving various machine learning models.