GoFormer: A GoLPP inspired transformer for functional brain graph learning and classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Graph learning is crucial for modeling complex data relationships.
• Traditional methods like graph-optimized locality preserving projections (GoLPP) adapt graph learning with dimensionality reduction.
• Transformers utilize self-attention for relationship modeling but require large datasets and lack inductive bias.

Purpose of the Study:

• To develop a novel method, GoFormer, combining the strengths of GoLPP and Transformer.
• To address the limitations of Transformers, such as weak inductive bias and data requirements, especially in medical applications.
• To improve the performance and interpretability of graph learning models.

Main Methods:

• Revisiting GoLPP's iterative process to mirror Transformer's self-attention mechanism.
• Designing GoFormer to integrate Transformer's sequence handling with GoLPP's parameter updating and sharing.
• Applying GoFormer to learn and classify brain graphs from functional magnetic resonance imaging (fMRI) data.

Main Results:

• GoFormer demonstrated superior performance compared to baseline and state-of-the-art methods in brain graph classification.
• The method effectively mitigates overfitting risks inherent in large-scale models.
• GoFormer offers enhanced interpretability, crucial for medical applications.

Conclusions:

• GoFormer successfully merges adaptive graph learning with self-attention mechanisms.
• The model provides a powerful and interpretable solution for analyzing complex graph data, particularly in medical diagnostics.
• GoFormer represents a significant advancement for applications requiring robust graph learning with limited data.