MTKSVCR: A novel multi-task multi-class support vector machine with safe acceleration rule.

Area of Science:

• Machine Learning
• Computational Science

Background:

• Regularized multi-task learning (RMTL) is effective for binary classification but suboptimal for multi-class problems due to information loss and class imbalance.
• Existing RMTL extensions for multi-class problems, like one-versus-one and one-versus-rest, do not fully leverage sample information.

Purpose of the Study:

• To propose an original multi-task multi-class model (MTKSVCR) using an "one-versus-one-versus-rest" strategy to enhance testing accuracy.
• To develop a safe acceleration (SA) rule to mitigate the computational time of MTKSVCR by reducing the optimization problem size.

Main Methods:

• MTKSVCR mines related information across multiple tasks by setting distinct penalty parameters for task-common and task-specific regularization terms.
• The SA rule identifies and removes superfluous samples with zero elements in the dual optimal solution before solving, reducing the problem size.
• The SA rule guarantees an identical optimal solution to the original problem, ensuring safety and effectiveness even with simultaneous parameter changes.

Main Results:

• Experiments on artificial and benchmark datasets demonstrate the effectiveness of the proposed MTKSVCR model.
• The SA rule significantly reduces the time consumption of the MTKSVCR model while maintaining solution accuracy.
• The proposed methods show validity and improved performance in multi-task multi-class learning scenarios.

Conclusions:

• MTKSVCR offers an improved approach for multi-task multi-class learning by fully utilizing sample information.
• The SA rule provides a safe and efficient method to accelerate the optimization process in MTKSVCR.
• The combined approach enhances computational efficiency and predictive accuracy in complex machine learning tasks.