Semantic Feature Learning for Heterogeneous Multitask Classification via Non-Negative Matrix Factorization.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Multitask learning (MTL) typically assumes uniform feature representations across tasks.
• This assumption is often violated in real-world scenarios, limiting MTL's effectiveness.
• Handling heterogeneous features in MTL is a significant challenge.

Purpose of the Study:

• To address multitask learning with heterogeneous features.
• To develop methods that can learn from tasks with diverse data representations.
• To improve generalization performance in complex, real-world applications.

Main Methods:

• Constructing an integrated graph to connect different tasks.
• Proposing Multitask Non-negative Matrix Factorization (MTNMF) to find a common semantic feature space.
• Introducing an improved MTNMF (IMTNMF) that avoids information loss by not requiring feature-label correlation matrices.

Main Results:

• The proposed MTNMF and IMTNMF methods effectively handle MTL with heterogeneous features.
• IMTNMF demonstrated improved performance over MTNMF, achieving approximately 2% higher average accuracy.
• The methods were validated through extensive experiments on three real-world problems.

Conclusions:

• MTL can be effectively performed even when tasks have heterogeneous features.
• The proposed graph-based and matrix factorization approaches provide a robust solution.
• IMTNMF offers a more accurate and efficient approach to heterogeneous MTL.