Convex formulation of multiple instance learning from positive and unlabeled bags

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Multiple Instance Learning (MIL) uses bag-level labels for data composed of instances.
• Traditional MIL requires fully labeled bags, which is often impractical.
• Positive and Unlabeled (PU) classification addresses scenarios with limited labeled data.

Purpose of the Study:

• To propose a novel convex PU classification method for MIL problems.
• To address the challenge of limited labeled bags in practical MIL applications.
• To improve efficiency and performance in PU-MIL.

Main Methods:

• Developed a convex optimization framework for PU classification within MIL.
• Integrated PU learning principles into the MIL framework.
• Conducted experimental evaluations to compare performance and computational cost.

Main Results:

• The proposed convex PU classification method demonstrated superior performance.
• Achieved significantly lower computational costs compared to existing methods.
• Validated effectiveness in addressing the challenge of unlabeled bags in MIL.

Conclusions:

• The novel convex PU classification method is effective for MIL.
• Offers a computationally efficient alternative for PU-MIL.
• Provides a practical solution for MIL problems with limited labeled data.