Graph embedded nonparametric mutual information for supervised dimensionality reduction.

Area of Science:

• Machine Learning
• Data Science
• Information Theory

Background:

• Dimensionality reduction is crucial for simplifying complex datasets.
• Mutual Information (MI) offers a robust criterion, correlating with Bayes error rate.
• Nonparametric MI implementations are computationally efficient and flexible.

Purpose of the Study:

• To develop a novel dimensionality reduction algorithm.
• To leverage mutual information (MI) as a criterion for data transformation.
• To integrate MI with graph embedding and linear methods.

Main Methods:

• Formulating nonparametric MI as a kernel objective within graph embedding.
• Developing a linear equivalent for efficient dimensionality reduction.
• Comparing proposed methods against state-of-the-art algorithms.

Main Results:

• The proposed nonparametric MI-based dimensionality reduction methods achieve competitive results.
• In most cases, the novel methods outperform existing dimensionality reduction techniques.
• Effectiveness validated across various classifiers and benchmark/real-life datasets.

Conclusions:

• Nonparametric mutual information is a powerful objective for dimensionality reduction.
• The developed algorithms offer an effective alternative to current methods.
• This approach enhances data analysis by preserving class-discriminative information.