Linear time relational prototype based learning.

Area of Science:

• Machine Learning
• Data Mining
• Bioinformatics

Background:

• Prototype-based learning provides intuitive data inspection in supervised and unsupervised settings.
• Existing methods for relational data (non-Euclidean) face quadratic time complexity, limiting scalability.
• This infeasibility hinders analysis of medium to large datasets.

Purpose of the Study:

• To develop a novel supervised prototype-based classification technique for dissimilarity data.
• To adapt linear time approximation techniques (Nyström approximation) for efficiency.
• To apply these methods to both supervised (LVQ-based) and unsupervised (GTM) learning for relational data.

Main Methods:

• Proposed a novel supervised prototype-based classification algorithm for dissimilarity data, extending Learning Vector Quantization (LVQ).
• Transferred the Nyström approximation technique to reduce quadratic time complexity.
• Applied the Nyström approximation to the proposed LVQ-based classifier and relational Generative Topographic Mapping (GTM).

Main Results:

• Achieved linear time and space complexity for prototype-based learning on relational data.
• Demonstrated the effectiveness of the novel techniques on three biomedical domain examples.
• Enabled efficient analysis of large-scale relational datasets previously computationally infeasible.

Conclusions:

• The developed methods offer computationally efficient solutions for prototype-based learning with relational data.
• These advancements are particularly relevant for the biomedical domain, facilitating analysis of complex datasets.
• The linear time complexity significantly enhances the scalability of these machine learning techniques.