Divergence-Based Locally Weighted Ensemble Clustering with Dictionary Learning and L_2,1-Norm.

Area of Science:

• Data Mining
• Machine Learning
• Artificial Intelligence

Background:

• Accurate clustering of unlabeled data remains a significant challenge.
• Ensemble clustering methods improve accuracy and stability by combining base clusterings.
• Existing methods like DREC and ELWEC have limitations in handling microcluster differences and sample-cluster relationships.

Purpose of the Study:

• To propose a novel ensemble clustering method, divergence-based locally weighted ensemble clustering with dictionary learning (DLWECDL).
• To address the limitations of existing ensemble clustering techniques by incorporating microcluster weighting and dictionary learning.
• To improve the accuracy and stability of clustering for unlabeled data.

Main Methods:

• Generating microclusters from base clustering results.
• Calculating microcluster weights using Kullback-Leibler divergence-based ensemble-driven cluster index.
• Employing an ensemble clustering algorithm with dictionary learning and L2,1-norm, optimizing via subproblems to learn a similarity matrix.
• Obtaining final clustering results by partitioning the similarity matrix using normalized cut (Ncut).

Main Results:

• The proposed DLWECDL method was validated on 20 diverse datasets.
• Experimental comparisons showed DLWECDL outperforming other state-of-the-art ensemble clustering methods.
• The results indicate DLWECDL's effectiveness in improving clustering accuracy.

Conclusions:

• DLWECDL offers a promising approach to ensemble clustering for unlabeled data.
• The method effectively addresses limitations of prior techniques by considering microcluster importance and sample-cluster relationships.
• DLWECDL demonstrates superior performance and stability in clustering tasks.