An efficient topological distance-based tree kernel.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Traditional tree kernels often overlook the relative positioning of substructures within trees.
• Existing methods primarily exploit structural information, neglecting orthogonal positional data.
• Combining these information types can potentially improve tree kernel accuracy.

Purpose of the Study:

• To develop an efficient algorithm for integrating positional information into subtree kernels.
• To enhance the feature space of tree kernels without increasing worst-case complexity.
• To evaluate the performance of the proposed method against state-of-the-art tree kernels.

Main Methods:

• An efficient algorithm was designed to inject positional information into tree kernels.
• Techniques were developed to expand the feature space of tree kernels.
• The proposed subtree kernel with positional information was implemented and tested.

Main Results:

• The enhanced tree kernel achieved state-of-the-art performance on several benchmark datasets.
• The method demonstrated superior or comparable performance to computationally intensive tree kernels.
• Positional information was effectively integrated, leading to improved accuracy.

Conclusions:

• Integrating positional information into tree kernels is a viable strategy for enhancing performance.
• The proposed algorithm offers an efficient way to leverage this orthogonal information.
• This approach advances the capabilities of tree-based machine learning models.