Robust Multiple Flat Projections Clustering With Truncated Distance Maximization Constraints.

Area of Science:

• Machine Learning
• Data Mining
• Clustering Algorithms

Background:

• Flat-type projection clustering methods explore multiple subspaces to enhance performance.
• Existing methods often use greedy search, leading to high computational costs and ignoring projection interdependencies.
• Current approaches struggle to effectively suppress outliers and noisy data, compromising cluster discrimination.

Purpose of the Study:

• To propose a robust multiple flat projections clustering (RMFPC) method.
• To address limitations of existing methods in handling outliers, noisy data, and computational efficiency.
• To improve data discrimination in flat-type projection clustering.

Main Methods:

• Utilizes the L2,1-norm for computing within-and between-cluster distances, enhancing outlier robustness.
• Introduces a truncated distance maximization constraint (TDMC) to mitigate the impact of noisy data on cluster separability.
• Develops an efficient non-greedy solution algorithm based on a novel, theoretically equivalent problem formulation.

Main Results:

• The proposed RMFPC method demonstrates enhanced robustness against outliers and noisy data.
• The efficient non-greedy algorithm achieves accurate cluster center estimation through an optimization mechanism.
• Experimental results on toy and real-world datasets validate the effectiveness of RMFPC.

Conclusions:

• RMFPC offers a superior approach to flat-type projection clustering compared to existing methods.
• The method effectively improves data discrimination by robustly handling outliers and noisy data.
• The developed algorithm is computationally efficient and theoretically sound, validated by empirical evidence.