Efficient feature selection and multiclass classification with integrated instance and model based learning.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Multiclass classification and feature selection are crucial in biological and medical fields.
• Existing methods like K nearest neighbor (KNN) struggle with high dimensionality, while logistic regression-based approaches face challenges with feature independence and data imbalance.
• Extending binary classification to multiclass problems presents significant difficulties.

Purpose of the Study:

• To develop an efficient learning method for simultaneous multiclass classification and feature selection.
• To address the limitations of existing instance-based and model-based methods in high-dimensional and unbalanced datasets.
• To propose a combined approach integrating KNN and constrained logistic regression (KNNLog).

Main Methods:

• Integration of instance-based K nearest neighbor (KNN) and model-based constrained logistic regression (KNNLog).
• Simultaneous minimization of intra-class distance and maximization of inter-class distance.
• Application to problems with small sample sizes and unbalanced classes, common in biological data.

Main Results:

• KNNLog demonstrates efficiency for small sample sizes and unbalanced classes.
• The method effectively identifies highly correlated features, mitigating issues from multiple testing.
• Evaluated on microRNA and metagenomic datasets, KNNLog showed strong performance.

Conclusions:

• KNNLog offers an efficient solution for simultaneous multiclass classification and feature selection.
• The proposed method is particularly suitable for challenging biological datasets with high dimensionality and imbalance.
• KNNLog provides a robust alternative to existing methods, improving feature selection and classification accuracy.