Efficient similarity-based data clustering by optimal object to cluster reallocation.

Area of Science:

• Machine Learning
• Data Mining
• Algorithm Design

Background:

• Clustering algorithms often rely on distance metrics, which may not fully capture the nuances of similarity data.
• Kernel k-means and spectral clustering are popular but have limitations regarding data types and computational efficiency.

Purpose of the Study:

• To introduce a novel iterative flat hard clustering algorithm.
• To optimize clustering by maximizing average intra-class similarity instead of minimizing squared distances.
• To enhance applicability to arbitrary symmetric similarity matrices and improve computational efficiency for large datasets.

Main Methods:

• Developed an iterative flat hard clustering algorithm operating on symmetric similarity matrices.
• Focused on maximizing average intra-class similarity as the core objective function.
• Evaluated performance against kernel k-means and spectral clustering on various datasets.

Main Results:

• The proposed algorithm achieves performance equivalent or superior to existing methods like kernel k-means and spectral clustering.
• Demonstrated significantly faster execution times compared to established algorithms.
• Showcased reduced memory access, proving advantageous for large-scale data analysis.

Conclusions:

• The novel similarity-based clustering approach offers a more effective and efficient alternative for data analysis.
• The algorithm's flexibility with similarity matrices and computational advantages make it suitable for big data challenges.
• Reproducibility is supported through publicly available materials.