Ensemble Clustering using Semidefinite Programming.

Area of Science:

• Data Science
• Machine Learning
• Computer Vision

Background:

• Ensemble clustering aims to aggregate multiple clustering solutions into a single, superior one.
• Existing methods often rely on voting strategies, which may not fully capture agreement.
• Maximizing shared information is key to effective ensemble clustering.

Purpose of the Study:

• To propose a novel agreement measure for ensemble clustering using 2D string encoding.
• To develop an optimization model for maximizing this new agreement measure.
• To demonstrate the effectiveness of the proposed method compared to existing strategies.

Main Methods:

• Introduced a 2D string encoding for measuring agreement between clusterings.
• Formulated a nonlinear optimization model to maximize the proposed agreement measure.
• Transformed the problem into a Semidefinite Program (SDP) solvable in polynomial time.

Main Results:

• The 2D string encoding provides a more robust measure of agreement.
• The proposed method shows improvements in both the new agreement measure and existing ones.
• Evaluations on clustering and image segmentation databases confirm the effectiveness.

Conclusions:

• The 2D string encoding offers a superior approach to ensemble clustering agreement.
• The developed SDP relaxation provides an efficient computational solution.
• This work advances ensemble clustering techniques for improved data analysis.