Multitask Feature Selection by Graph-Clustered Feature Sharing.

Area of Science:

• Machine Learning
• Data Science
• Computational Biology

Background:

• Multitask feature selection (MTFS) is crucial for high-dimensional data.
• Existing methods often assume shared features across all tasks, which can lead to negative transfer if task correlations are weak.
• This limitation hinders performance when task relationships are not clearly defined.

Purpose of the Study:

• To develop a flexible MTFS method that accommodates varying task correlations.
• To address the limitations of traditional MTFS by avoiding rigid task partitioning.
• To improve feature selection accuracy in multitask learning scenarios.

Main Methods:

• A graph-clustered feature sharing approach is proposed to represent task relevance.
• Graph-guided regularization is introduced for sparsity at both task and feature levels.
• A smooth proximal gradient method is employed to solve the optimization problem.

Main Results:

• The proposed method effectively captures underlying task structures.
• Evaluations on multitask regression and binary classification problems show strong performance.
• Experiments on synthetic and real-world datasets validate the approach's effectiveness.

Conclusions:

• The graph-clustered MTFS method offers a flexible and effective solution for high-dimensional multitask learning.
• It successfully mitigates negative transfer by modeling task relationships via graphs.
• The approach demonstrates significant potential for various real-world applications requiring robust feature selection.