Multilinear Multitask Learning by Rank-Product Regularization.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Traditional multitask learning (MTL) captures first-order correlations between tasks.
• Multilinear multitask learning (MLMTL) extends MTL by considering higher-order correlations among tasks arranged by multiple indices.
• Existing MLMTL methods often rely on matrix-based regularization, limiting the exploitation of tensor structures.

Purpose of the Study:

• To propose a novel Multilinear Multitask Learning (MLMTL) model that effectively utilizes higher-order correlations among tasks.
• To introduce a rank-product regularization term designed to capture the latent correlation structure within a coefficient tensor.
• To improve the overall performance, accuracy, and efficiency of multitask learning by better describing intrinsic high-order correlations.

Main Methods:

• Developed a new MLMTL model incorporating a rank-product regularization term into the objective function.
• Designed an efficient optimization algorithm using the alternating direction method of multipliers (ADMM) to solve the proposed model.
• Analyzed the convergence properties of the ADMM algorithm, demonstrating its asymptotic regularity under certain conditions.

Main Results:

• The proposed MLMTL model with rank-product regularization effectively captures higher-order correlations among tasks.
• Experiments on synthetic and real-world datasets show superior performance compared to existing MLMTL methods.
• The developed ADMM algorithm provides an efficient and convergent solution for the optimization problem.

Conclusions:

• The novel MLMTL approach with rank-product regularization offers a more precise way to describe high-order task correlations.
• This method leads to significant improvements in accuracy and efficiency for multitask learning problems.
• The proposed approach advances the field of MLMTL by providing a robust and effective framework.