Subgraph-Mamba: Subgraph Mamba model with positional encoding.

Area of Science:

• Graph Representation Learning
• Machine Learning
• Artificial Intelligence

Background:

• Subgraph representation learning is vital but faces challenges with node dependencies and feature inaccuracies due to subgraph overlaps.
• Existing methods struggle to effectively model relationships within and between subgraphs.

Purpose of the Study:

• To propose a novel method, Subgraph Mamba Model with Positional Encoding (Subgraph-Mamba), to address limitations in current subgraph representation learning techniques.
• To enhance the capture of dependencies among subgraph nodes and mitigate feature inaccuracies.

Main Methods:

• Employing Graph Convolution with Different Weights (GDW) for feature extraction and updates.
• Introducing Sine Projection Positional Encoding (SPPE) to encode subgraph positions and integrate them into node features.
• Utilizing a Multi-head Mamba module to capture complex dependencies within subgraph node features.

Main Results:

• Subgraph-Mamba successfully incorporates positional encoding into subgraph node features.
• The Multi-head Mamba module effectively captures dependencies among subgraph node features.
• Subgraph-Mamba achieves superior subgraph feature representation by aggregating node features and dependencies.

Conclusions:

• Subgraph-Mamba demonstrates superior performance compared to state-of-the-art baselines in subgraph representation learning.
• This work represents the first application of Mamba architecture to subgraph representation learning.
• The proposed method offers a promising direction for future research in graph representation learning.