GraphHop: An Enhanced Label Propagation Method for Node Classification.

Area of Science:

• Graph Machine Learning
• Network Science
• Data Mining

Background:

• Semisupervised node classification on graph-structured data is crucial for many applications.
• Graph convolutional networks (GCNs) show promise but struggle with scalability and large datasets.
• Existing methods face challenges with high model complexity and the need for extensive labeled data.

Purpose of the Study:

• To propose GraphHop, a scalable semisupervised node classification method.
• To address limitations of current GCNs in handling large-scale graphs and complex optimization.
• To improve label propagation (LP) methods for enhanced graph learning.

Main Methods:

• GraphHop employs an enhanced label propagation algorithm with alternating label aggregation and update steps.
• Multihop neighbor embeddings are aggregated to center nodes for improved signal smoothing.
• A two-stage training process (initialization and iteration) effectively encodes attributes, links, and labels.

Main Results:

• GraphHop demonstrates superior performance compared to state-of-the-art graph learning methods.
• The method achieves strong results across diverse tasks, including multilabel and multiclass classification.
• Effectiveness is validated on citation networks, social graphs, and commodity consumption graphs of various sizes.

Conclusions:

• GraphHop offers a scalable and effective solution for semisupervised node classification on large graphs.
• The enhanced LP approach improves graph signal smoothing and learning capacity.
• GraphHop provides a robust framework for encoding complex graph information, outperforming existing techniques.