Context-Dependent Random Walk Graph Kernels and Tree Pattern Graph Matching Kernels with Applications to Action...

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Graphs are essential for modeling complex data structures.
• Existing graph analysis methods may not fully capture intricate relationships.

Purpose of the Study:

• To propose new families of graph kernels for enhanced data analysis.
• To improve graph similarity measurement and discriminative power for pattern recognition tasks.

Main Methods:

• Developed context-dependent random walk graph kernels incorporating contextual information.
• Introduced tree-pattern graph matching kernels utilizing quadratic optimization with sparse constraints.
• Applied multiple kernel learning with l1,2-norm regularization to combine kernels.

Main Results:

• The proposed kernels effectively measure graph similarity and enhance discriminative capabilities.
• Applied to human action recognition, the kernels demonstrated superior performance compared to state-of-the-art methods.
• Tree-pattern kernels achieved higher accuracy but with increased computational cost compared to random walk kernels.

Conclusions:

• The novel graph kernels offer significant improvements for complex data modeling and recognition tasks.
• Context-dependent random walk and tree-pattern matching kernels provide effective solutions for human action recognition.