Attribute Augmented Network Embedding Based on Generative Adversarial Nets.

Area of Science:

• Computer Science
• Data Science
• Network Analysis

Background:

• Network embedding aims to create low-dimensional node representations for network analysis.
• Existing methods struggle with heterogeneous structure and attribute data, potentially leading to overfitting.
• Attribute augmented networks offer a unified framework for integrating diverse network information.

Purpose of the Study:

• To develop a novel network embedding technique that effectively handles networks with attribute information.
• To leverage generative adversarial nets (ANGANs) for learning robust and informative node representations.
• To address the limitations of existing methods in preserving both structural and attribute features.

Main Methods:

• An attribute augmented network was employed to unify structure and attribute information.
• Generative adversarial nets (ANGANs) were utilized for adversarial learning between generative and discriminative models.
• The generative model enhances the Skip-gram model, while the discriminative model acts as a binary classifier; pre-training and teacher forcing improve stability.

Main Results:

• The proposed ANGAN method demonstrated superior performance compared to state-of-the-art techniques.
• Empirical results confirmed the effectiveness and generality of the ANGAN approach across various applications.
• The method successfully learned low-dimensional representations capturing both network structure and node attributes.

Conclusions:

• ANGAN provides an effective solution for attribute augmented network embedding.
• The method offers robust and informative node representations, outperforming existing approaches.
• ANGAN shows significant potential for diverse real-world network analysis tasks.