Local and global preserving semisupervised dimensionality reduction based on random subspace for cancer...

Area of Science:

• Bioinformatics
• Machine Learning
• Cancer Research

Background:

• Accurate cancer classification is crucial for effective diagnosis and treatment.
• Existing semisupervised dimensionality reduction methods struggle with noisy, high-dimensional data, leading to unstable neighborhood topology.
• This instability hinders the performance of classification algorithms on complex biological datasets.

Purpose of the Study:

• To propose a novel semisupervised dimensionality reduction algorithm, RSLGSSDR (Random Subspace Local and Global preserving Semisupervised Dimensionality Reduction).
• To address the instability of neighborhood graph construction in high-dimensional, noisy datasets.
• To enhance the accuracy and robustness of cancer classification using gene expression data.

Main Methods:

• Developed RSLGSSDR, a semisupervised dimensionality reduction technique utilizing the random subspace algorithm.
• Constructed multiple diverse graphs across random subspaces of the dataset.
• Fused these graphs into a mixture graph for dimensionality reduction, enabling robust graph construction in lower dimensions.

Main Results:

• RSLGSSDR demonstrated superior recognition performance compared to existing competitive methods on public gene expression datasets.
• The algorithm proved to be robust across a wide range of input parameter values.
• The mixture graph approach effectively handles complex geometric distributions in high-dimensional data.

Conclusions:

• RSLGSSDR offers a robust and effective solution for semisupervised dimensionality reduction in cancer classification.
• The method overcomes limitations of traditional approaches when dealing with noisy, high-dimensional biological data.
• This advancement holds promise for improving cancer diagnosis and treatment strategies through enhanced classification accuracy.