Multigranularity Fuzzy Autoencoder for Discriminative Feature Selection in High-Dimensional Data.

Area of Science:

• Bioinformatics
• Machine Learning
• Computational Biology

Background:

• Biological datasets, like gene expression data, are often high-dimensional, leading to overfitting and computational challenges.
• Traditional feature selection methods struggle with noisy data, diminishing the performance of deep learning models like autoencoders.
• Effective feature selection is crucial for reducing dimensionality, improving model performance, and enhancing data interpretability.

Purpose of the Study:

• To propose a novel feature selection method, multigranularity fuzzy autoencoders (FAEs), to address challenges in high-dimensional and noisy biological data.
• To enhance the management of noise and outliers in biological datasets through the integration of fuzzy theory and autoencoder models.
• To improve classification accuracy and robustness by developing a more effective feature selection technique.

Main Methods:

• Developed multigranularity fuzzy autoencoders (FAEs) integrating fuzzy theory with autoencoder models.
• Introduced a feature selection layer approximating discrete selection using continuous probability distributions.
• Incorporated a coarse-grained loss function for discriminative power and intuitionistic fuzzy weights to manage uncertainty and mitigate noise.

Main Results:

• Demonstrated significant improvements in feature selection effectiveness across 20 public datasets.
• Validated the approach on a real-world schizophrenia dataset, showing enhanced classification accuracy and robustness.
• Outperformed existing feature selection techniques in managing noisy and high-dimensional biological data.

Conclusions:

• The proposed FAE method offers a robust solution for feature selection in high-dimensional and noisy biological data.
• This approach shows significant potential for improving classification performance in various biological research areas, including schizophrenia.
• FAEs provide a promising direction for advancing machine learning applications in bioinformatics and computational biology.