GRAPE for fast and scalable graph processing and random-walk-based embedding.

Area of Science:

• Computer Science
• Data Science
• Machine Learning

Background:

• Graph representation learning methods are crucial for complex real-world problems.
• Current methods struggle with large-scale graphs (millions of nodes, billions of edges).

Purpose of the Study:

• Introduce GRAPE (Graph Representation Learning, Prediction and Evaluation) as a scalable software resource.
• Address the limitations of existing software for big graph processing and embedding.

Main Methods:

• Developed GRAPE using specialized data structures, algorithms, and parallel implementation of random-walk-based methods.
• Implemented GRAPE in Python and Rust, comprising 1.7 million lines of code.
• Integrated 69 node-embedding methods, 25 inference models, and efficient graph-processing utilities.

Main Results:

• GRAPE demonstrates orders of magnitude improvement in space and time complexity compared to state-of-the-art.
• Achieved competitive performance in edge- and node-label prediction.
• Provides over 80,000 graphs for diverse applications.

Conclusions:

• GRAPE offers a scalable solution for big graph analysis and representation learning.
• Standardized interfaces and modular pipelines facilitate integration and fair comparison of methods.
• Positions GRAPE as a valuable resource for the graph machine learning community.