Unsupervised Graph Embedding via Adaptive Graph Learning.

Area of Science:

• Machine Learning
• Graph Representation Learning

Background:

• Graph autoencoders (GAEs) are effective for representation learning but are sensitive to the quality of graph structures (adjacency matrices).
• Incomplete or noisy adjacency matrices can significantly degrade GAE performance.

Purpose of the Study:

• To propose novel unsupervised graph embedding methods, BAGE and VBAGE, that are robust to incomplete or disturbed graph structures.
• To expand the applicability of GAEs to datasets lacking explicit graph structures.

Main Methods:

• Introduced unsupervised graph embedding via adaptive graph learning (BAGE) and variational adaptive graph learning (VBAGE).
• Developed an adaptive learning mechanism to initialize and self-learn the adjacency matrix, independent of initial parameters.
• Embedded latent representations with Laplacian graph structure to preserve topological information.

Main Results:

• The proposed methods demonstrate robustness to graph structure imperfections.
• BAGE and VBAGE significantly outperform baseline methods across multiple tasks.
• Validated through extensive experiments on various datasets.

Conclusions:

• The adaptive learning approach enhances the robustness and applicability of graph autoencoders.
• BAGE and VBAGE offer superior performance in node clustering, classification, link prediction, and graph visualization.
• These methods provide a powerful solution for graph embedding on incomplete or general datasets.