Weighted Mutual Information for Aggregated Kernel Clustering.

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• Clustering is essential in machine learning for grouping similar data.
• Linear methods like K-means and non-linear kernel K-means exist.
• Kernel K-means uses kernel functions for non-linear data projection, but kernel choice is dataset-dependent.

Purpose of the Study:

• To develop an impartial clustering method that aggregates results from different kernels.
• To address the challenge of selecting the optimal kernel for arbitrary datasets.
• To introduce a novel weighting approach for combining clustering outcomes.

Main Methods:

• Proposed Weighted Mutual Information (WMI) for calculating weights.
• Weights are based on the performance of individual clustering methods on a labeled training set.
• Aggregated results using a weight function to combine individual clustering outcomes.

Main Results:

• Applied WMI to four non-linearly separable datasets.
• Evaluated performance under various noise conditions.
• Demonstrated WMI's effectiveness in improving clustering accuracy.

Conclusions:

• Weighted Mutual Information provides an impartial clustering solution.
• The method is independent of a single kernel choice.
• WMI outperforms individual kernel K-means methods, particularly in high-noise environments.