Reinforced GNNs for Multiple Instance Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence

Background:

• Multiple instance learning (MIL) trains models using bags of instances with only bag-level labels.
• Graph neural networks (GNNs) enhance MIL by capturing intrabag topology.
• Current GNNs require manual, asynchronous adjustments for graph structures and architectures, which is inefficient.

Purpose of the Study:

• To propose a novel reinforced GNN framework for MIL (RGMIL).
• To address the limitations of manual and asynchronous adjustments in existing GNNs for MIL.
• To leverage multiagent deep reinforcement learning (MADRL) for automated and synchronized control.

Main Methods:

• Developed a reinforced GNN framework for MIL (RGMIL).
• Pioneered the use of multiagent deep reinforcement learning (MADRL) in MIL tasks.
• Enabled flexible definition of factors influencing bag graphs and GNNs, with synchronous control.

Main Results:

• RGMIL automates adjustments by exploring structure-to-architecture correlations.
• Experimental results on multiple MIL datasets show RGMIL achieves superior performance.
• The framework demonstrates excellent explainability.

Conclusions:

• RGMIL offers an effective and automated solution for MIL using GNNs and MADRL.
• The synchronized control of graph structure and GNN architecture improves learning efficiency.
• The proposed method sets a new benchmark for performance and explainability in MIL.