Graph Learning on Millions of Data in Seconds: Label Propagation Acceleration on Graph Using Data Distribution.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Graph-based semi-supervised learning is vital for applications like social mining and multimedia classification.
• Existing methods struggle with large-scale datasets due to high computational complexity and lack of incremental learning capabilities.
• Scalability and adaptability are crucial for real-world data that continuously grows.

Purpose of the Study:

• To propose a novel method, Data Distribution Based Graph Learning (DDGL), for efficient semi-supervised learning on large-scale datasets.
• To address the limitations of existing methods regarding computational complexity and incremental learning.
• To enhance label propagation speed and prediction accuracy in dynamic, large-scale environments.

Main Methods:

• DDGL propagates labels using smaller-scale data distribution model parameters, bypassing direct computation on raw data.
• An adaptive graph updating strategy is employed to manage distribution bias between new and existing data, facilitating incremental learning.
• The method was validated through comprehensive experiments on datasets ranging from seven thousand to five million samples.

Main Results:

• DDGL significantly accelerates label propagation compared to traditional methods.
• The proposed method demonstrates improved prediction accuracy by preserving structural information.
• Experiments show substantial improvements in classification accuracy and reduced computation time on large-scale datasets.

Conclusions:

• DDGL offers a computationally efficient and accurate solution for semi-supervised learning on large-scale, evolving datasets.
• The method's ability to support incremental learning makes it suitable for real-world applications with continuously increasing data volumes.
• DDGL represents a significant advancement over existing graph-based semi-supervised learning techniques.