Fusion of text and graph information for machine learning problems on networks

Area of Science:

• Computer Science
• Data Science
• Machine Learning

Background:

• Network representation learning maps network nodes into low-dimensional vectors.
• These embeddings preserve network properties and aid downstream tasks like node classification.
• Networks often contain associated text, such as scientific paper abstracts or social media posts.

Purpose of the Study:

• To explore combining existing text and network embedding methods.
• To enhance accuracy for downstream machine learning tasks.
• To propose modifications for better textual information capture in network embedding frameworks.

Main Methods:

• Investigated fusion of text embedding techniques with network embedding architectures.
• Proposed modifications to popular network embedding models.
• Evaluated performance on downstream machine learning tasks.

Main Results:

• Combining text and network embeddings improved accuracy for downstream tasks.
• Modified architectures better captured textual information within network embeddings.
• Demonstrated the benefit of multimodal data integration.

Conclusions:

• Integrating textual information significantly enhances network representation learning.
• Proposed modifications offer improved methods for multimodal network analysis.
• This approach advances capabilities in tasks requiring both network structure and content understanding.