Genetic Programming for Evolving Similarity Functions for Clustering: Representations and Analysis.

Area of Science:

• Data Mining
• Machine Learning
• Artificial Intelligence

Background:

• Clustering algorithms commonly rely on predefined similarity measures like Euclidean distance.
• These fixed measures limit adaptability and interpretability for diverse datasets.
• Tailoring similarity functions to dataset properties is crucial for improved clustering.

Purpose of the Study:

• To develop a novel approach for automatically evolving similarity functions for clustering.
• To enhance clustering performance and interpretability through dataset-specific similarity measures.
• To leverage genetic programming for feature selection and construction in similarity function evolution.

Main Methods:

• Utilized genetic programming to automatically evolve similarity functions.
• Implemented a method combining feature selection and feature construction.
• Employed a graph-based representation for clustering with evolved similarity functions.
• Extended the approach using a multi-tree method for complementary similarity functions.

Main Results:

• Demonstrated superior and more consistent performance compared to benchmark methods on large, high-dimensional datasets.
• Showcased the effectiveness of evolved similarity functions in improving clustering quality.
• Achieved further performance gains through the multi-tree approach for complementary functions.
• Analyzed the interpretability and structure of evolved functions, providing insights into their superiority.

Conclusions:

• The proposed genetic programming-based method effectively evolves dynamic and flexible similarity functions for clustering.
• Dataset-specific similarity functions significantly outperform standard distance metrics in terms of performance and interpretability.
• The approach offers a powerful tool for advancing data mining and machine learning applications.