Improving Embedding of Graphs With Missing Data by Soft Manifolds.

Area of Science:

• Graph theory
• Computational geometry
• Machine learning

Background:

• Graph embeddings are crucial for automatic information extraction.
• Current methods assume Euclidean tangent spaces, limiting their use with complex, real-world graphs.
• Real-world graphs often have weighted connections and sparse, incomplete data.

Purpose of the Study:

• To introduce a novel manifold structure, the soft manifold, to address limitations in current graph embedding techniques.
• To enable more accurate and reliable characterization of complex graphs in continuous spaces.

Main Methods:

• Developed a new class of manifold named 'soft manifold' with spherical symmetry.
• Tangent spaces in soft manifolds are hypocycloids, shaped by information propagation velocity.
• Applied the soft manifold approach to reconstruction tasks on synthetic and real datasets.

Main Results:

• The soft manifold approach demonstrated more accurate graph characterization compared to state-of-the-art methods.
• Experimental results showed improved reliability in embedding complex graph structures.
• The method effectively handles weighted connections and sparse data inherent in real-world graphs.

Conclusions:

• Soft manifolds offer a robust solution for embedding complex graphs in continuous spaces.
• This advancement enhances the reliability of information extraction from diverse graph datasets.
• The proposed method represents a significant improvement over existing manifold-based graph embedding techniques.