Multitask Feature Learning Meets Robust Tensor Decomposition for EEG Classification.

Area of Science:

• Machine Learning
• Computer Science
• Biomedical Engineering

Background:

• Limited training samples are common in real-world classification tasks.
• Multitask learning (MTL) leverages shared information across related tasks for improved generalization.
• Tensors naturally represent multidimensional data common in various applications.

Purpose of the Study:

• To propose a regularized tensor-based multitask learning (MTL) method for joint feature selection and classification.
• To address challenges posed by limited training data in complex classification problems.
• To improve classification performance by effectively utilizing shared and task-specific information.

Main Methods:

• A tensor-based multitask learning framework is developed for classification.
• Fisher discriminant criterion is employed for discriminative feature selection and controlling within-class nonstationarity.
• A composite tensor norm, utilizing scaled latent trace norm and l1-norm, is proposed for regularizing shared and task-specific tensors.
• An efficient optimization algorithm based on Alternating Direction Method of Multipliers (ADMMs) is utilized.

Main Results:

• The proposed method effectively performs joint feature selection and classification.
• The method demonstrates superior performance compared to existing state-of-the-art techniques.
• Experimental validation on real electroencephalography (EEG) datasets confirms the method's efficacy.

Conclusions:

• The developed tensor-based MTL method offers a powerful approach for classification with limited data.
• The integration of feature selection and classification within a unified framework enhances model performance.
• The method shows significant potential for applications in areas like biomedical signal processing.