Optimizing Evaluation Metrics for Multitask Learning via the Alternating Direction Method of Multipliers

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Multitask learning (MTL) leverages shared factors across tasks to improve generalization.
• Current MTL methods often minimize errors (classification/regression) rather than direct evaluation metrics.
• Existing performance metrics include F1-score and area under the ROC curve.

Purpose of the Study:

• To propose a novel multitask learning (MTL) method that directly optimizes evaluation metrics.
• To address the optimization challenges posed by non-smooth components in the proposed formulation.
• To enhance the generalization performance of MTL across a wide range of problems.

Main Methods:

• Developed an MTL formulation combining a regularizer for task relatedness and structured hinge losses for metric optimization.
• Proposed a novel optimization procedure using the alternating direction method of multipliers (ADMM).
• Decomposed the problem into subproblems solvable via closed-form solutions and an efficient primal-dual algorithm.

Main Results:

• The proposed MTL method demonstrated superior performance gains compared to baseline methods.
• Extensive evaluations confirmed the effectiveness of directly optimizing evaluation metrics.
• The method showed significant improvements in a large family of MTL problems.

Conclusions:

• Directly optimizing evaluation metrics offers a more effective approach for multitask learning.
• The proposed ADMM-based optimization procedure efficiently handles non-smooth formulations.
• This work advances MTL by providing a method that directly targets desired performance measures.