Significant subgraph mining for neural network inference with multiple comparisons correction.

Area of Science:

• Computational neuroscience
• Graph theory
• Machine learning

Background:

• Neural network comparison is crucial for understanding complex systems.
• Existing methods may not fully capture differences in graph-generating processes.
• Significant subgraph mining presents a novel computational approach.

Purpose of the Study:

• To introduce and extend significant subgraph mining for comparing sets of unweighted graphs.
• To evaluate the method's statistical properties and provide practical recommendations for neuroscience applications.
• To apply the method to analyze differences in brain networks between patient groups.

Main Methods:

• Application of significant subgraph mining to compare graph sets.
• Extension of the method for dependent graph generating processes.
• Simulation studies using Erdős-Rényi models and empirical analysis of resting-state MEG data.
• Power analysis for transfer entropy networks.

Main Results:

• Demonstrated utility of significant subgraph mining in neural network comparison.
• Provided an extension for dependent graph processes, relevant for within-subject designs.
• Derived practical recommendations for applying subgraph mining in neuroscience through extensive error-statistical investigations.
• Empirical power analysis showed the method's effectiveness in distinguishing networks in autism spectrum disorder.

Conclusions:

• Significant subgraph mining is a valuable tool for comparing neural networks and understanding generative processes.
• The extended method and practical guidelines facilitate its application in neuroscience research.
• The Python implementation in the IDTxl toolbox enhances accessibility for researchers.