Survey on graph embeddings and their applications to machine learning problems on graphs.

Area of Science:

• Computer Science
• Machine Learning
• Data Science

Background:

• Relational data analysis traditionally demands extensive computational resources, domain expertise, and manual feature engineering.
• Automated graph feature engineering techniques, particularly graph embeddings, have emerged to address these challenges.
• Graph embeddings create vectorized feature spaces for graph components, enabling standard machine learning frameworks.

Purpose of the Study:

• To survey core concepts and taxonomies of graph embedding methodologies.
• To analyze the impact of network types on embedding structural and attributed data.
• To evaluate graph embedding applications in machine learning tasks on graphs.

Main Methods:

• Categorization of graph embedding models into matrix factorization, random-walks, and deep learning approaches.
• Analysis of how network characteristics influence the incorporation of structural and attributed data.
• Comprehensive evaluation of graph embedding applications including node classification, link prediction, clustering, and visualization.

Main Results:

• Identified three primary methodological approaches for graph embeddings.
• Demonstrated the influence of network types on embedding capabilities for diverse data.
• Provided experimental results on the quality of graph embeddings for key machine learning tasks.

Conclusions:

• Graph embeddings represent a significant advancement in network feature engineering.
• The survey offers an in-depth analysis of models and their applications across various machine learning problems on graphs.
• This work highlights open problems and future research directions in the rapidly growing field of graph embeddings.