Scalably Using Node Attributes and Graph Structure for Node Classification

Area of Science:

• Graph Machine Learning
• Network Science
• Data Mining

Background:

• Node classification infers unknown labels in networks using known labels and attributes.
• Traditional methods assume attribute similarity between adjacent nodes, which may not always hold.
• Existing approaches may fail when nodes with similar attributes are not adjacent or vice-versa.

Purpose of the Study:

• To introduce JANE (Jointly using Attributes and Node Embeddings), a flexible node classification approach.
• To address limitations of benchmark algorithms in diverse network settings.
• To provide a scalable and accurate solution for node classification.

Main Methods:

• JANE jointly leverages known node labels, network structure, and node attributes.
• The approach adapts to scenarios where labels are predicted from neighbors, attributes, or both.
• Experiments were conducted on synthetic data and seven real-world datasets of varying sizes and homophily.

Main Results:

• JANE demonstrates versatility and overcomes limitations of existing benchmark algorithms on synthetic data.
• The method scales effectively to large networks (up to 1.5M nodes).
• JANE achieves up to a 20% improvement in classification accuracy compared to strong baselines on real datasets.

Conclusions:

• JANE offers a principled and adaptable framework for node classification.
• The method provides significant accuracy improvements and scalability for real-world network analysis.
• JANE is effective across a wide range of network structures and attribute correlations.