Multitask Learning for Classification Problem via New Tight Relaxation of Rank Minimization

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Multitask learning (MTL) enhances performance by jointly training related tasks.
• Identifying shared low-dimensional latent subspaces across tasks is crucial for MTL.
• Trace norm is a common but not always tight approximation for rank minimization in MTL.

Purpose of the Study:

• To propose novel regularization-based models for MTL that offer tighter approximations of rank minimization.
• To improve the capture of low-dimensional latent subspaces shared across multiple tasks.
• To develop effective optimization strategies for the proposed NP-hard rank minimization problem.

Main Methods:

• Developed two novel regularization-based models minimizing the k minimal singular values.
• Proposed optimization strategies involving large penalizing parameters to solve the NP-hard rank minimization problem.
• Evaluated models on synthetic and real-world benchmark datasets.

Main Results:

• The proposed models provide tighter approximations to rank minimization compared to standard trace norm.
• The new models effectively capture the low-rank structure shared across tasks.
• Experimental results show superior performance over classical MTL methods.

Conclusions:

• The novel regularization models offer a more accurate approach to identifying shared latent structures in MTL.
• These models demonstrate improved performance in capturing task relatedness and overall predictive accuracy.
• The proposed optimization strategies successfully address the computational challenges of the rank minimization problem.