TACN: A Topical Adversarial Capsule Network for textual network embedding.

Area of Science:

• Computer Science
• Network Science
• Artificial Intelligence

Background:

• Effective network embedding requires integrating topological and attributed information.
• Prior methods often overlook or oversimplify node attributes, especially textual data, missing crucial latent topic information.
• Textual networks are increasingly prevalent, making the extraction of topic relevance vital for network analysis.

Purpose of the Study:

• To develop a novel network embedding model, the Topical Adversarial Capsule Network (TACN), for textual networks.
• To effectively integrate node structure, vertex attributes, and latent topic information from text into a unified low-dimensional representation.
• To improve network analysis by leveraging hidden information within textual attributes.

Main Methods:

• The proposed TACN model comprises three parts: an embedding model, a prediction model, and an adversarial capsule model.
• The embedding model extracts representations from network topology, vertex attributes, and document-topic distributions generated via a neural topic model.
• An adversarial capsule model is employed to fuse information from these diverse domains, incorporating adversarial principles for enhanced representation learning.

Main Results:

• Experiments on seven real-world datasets demonstrate the effectiveness of the TACN model.
• The TACN model successfully extracts low-dimensional latent spaces by combining structural, attribute, and topic information.
• The method shows superior performance in network embedding tasks involving textual data.

Conclusions:

• The Topical Adversarial Capsule Network (TACN) provides a robust framework for textual network embedding.
• Integrating topological, attribute, and latent topic information significantly enhances network representation.
• The TACN model offers a promising approach for analyzing complex textual networks.