Graph Attention Interaction Aggregation Network for Click-Through Rate Prediction.

Area of Science:

• Computational advertising
• Machine learning
• Graph neural networks

Background:

• Click-through rate (CTR) prediction is crucial for online advertising and recommendation systems.
• Current models often use implicit feature interactions, leading to poor interpretability and potential noise.
• This limits the predictive performance and understanding of complex feature relationships.

Purpose of the Study:

• To propose a novel CTR prediction model, GAIAN (Graph Attention Interactive Aggregation Network).
• To explicitly model feature interactions using a graph structure for enhanced interpretability and accuracy.
• To address limitations of implicit interaction modeling in existing CTR prediction methods.

Main Methods:

• GAIAN utilizes a graph attention interactive aggregation network to explicitly capture cross-feature interactions.
• A feature interactive selection mechanism is employed to identify beneficial cross features, reducing model noise.
• Bilinear interaction functions are integrated into the graph neural network aggregation strategy for fine-grained feature extraction.

Main Results:

• GAIAN demonstrates superior performance compared to state-of-the-art models on the Criteo and Avazu datasets.
• The model effectively extracts fine-grained intersection features in an explicit and flexible manner.
• Experimental results validate the model's ability to reduce noise and mitigate overfitting risks.

Conclusions:

• GAIAN enhances CTR prediction by explicitly modeling feature interactions through a graph attention network.
• The proposed method improves model interpretability and predictive accuracy.
• GAIAN offers a promising approach for computational advertising and recommendation systems.