Reverse Graph Learning for Graph Neural Network.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Data Mining

Background:

• Graph Neural Networks (GNNs) excel at feature learning by preserving local data structure.
• Existing GNNs struggle with noisy initial graphs (faulty/missing edges) and out-of-sample prediction.
• These limitations hinder performance in real-world applications.

Purpose of the Study:

• To propose a reverse GNN model for learning graphs from intrinsic data spaces.
• To develop a new out-of-sample extension method for GNNs.
• To improve feature learning quality and enable supervised/semi-supervised learning on unseen data.

Main Methods:

• A reverse GNN model is proposed to learn an optimal graph structure.
• An out-of-sample extension technique is introduced to handle unseen data points.
• The model integrates graph learning and prediction for enhanced GNN capabilities.

Main Results:

• The reverse GNN model generates high-quality graphs, improving feature learning.
• The out-of-sample extension method allows for effective supervised and semi-supervised learning.
• Competitive performance achieved in node classification, link prediction, and image retrieval tasks.

Conclusions:

• The proposed reverse GNN approach effectively addresses limitations of traditional GNNs.
• This method offers a robust solution for feature learning and out-of-sample prediction.
• The approach demonstrates strong potential for various graph-based machine learning applications.