Multitarget Sparse Latent Regression.

Area of Science:

• Machine Learning
• Statistical Modeling
• Data Science

Background:

• Multitarget regression is gaining popularity for solving multiple regression tasks simultaneously.
• Challenges include jointly modeling inter-target correlations and input-output relationships.

Purpose of the Study:

• To propose a novel framework, multitarget sparse latent regression (MSLR), for simultaneous multitarget regression.
• To effectively model intrinsic inter-target correlations and complex nonlinear input-output relationships.

Main Methods:

• MSLR employs a structure matrix to encode inter-target correlations via L-norm-based sparse learning.
• Kernel extension via a representer theorem allows handling complex nonlinear input-output relationships.
• An alternating optimization algorithm ensures efficient computation with guaranteed convergence.

Main Results:

• MSLR consistently outperforms state-of-the-art algorithms on synthetic and diverse real-world datasets.
• The framework demonstrates effectiveness in multivariate prediction.

Conclusions:

• MSLR provides an effective single framework for multitarget regression.
• The proposed method successfully addresses challenges in modeling inter-target correlations and nonlinearities.