Heterogeneous Multitask Metric Learning Across Multiple Domains.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Transfer metric learning (TML) leverages related domains for limited target data.
• Current TML often assumes homogeneous feature representations across domains.
• Existing heterogeneous transfer learning methods struggle with multiple domains.

Purpose of the Study:

• To develop a novel heterogeneous multitask metric learning (HMTML) framework.
• To enable learning metrics across multiple heterogeneous domains simultaneously.
• To address limitations of existing methods in handling multiple heterogeneous domains and high-order statistics.

Main Methods:

• The HMTML framework learns metrics for all domains together.
• Learned transformations induce a common subspace.
• High-order covariance among predictive structures is maximized within this subspace.

Main Results:

• HMTML effectively explores high-order statistical information across multiple domains.
• The proposed method yields more reliable feature transformations and learned metrics.
• Validated effectiveness on text categorization, scene classification, and social image annotation tasks.

Conclusions:

• HMTML provides a robust solution for metric learning in multiple heterogeneous domains.
• The framework's ability to exploit high-order statistics enhances transfer learning performance.
• This approach advances heterogeneous transfer learning capabilities for complex real-world applications.