Positive and Unlabeled Multi-Graph Learning.

Area of Science:

• Machine Learning
• Computer Science
• Data Science

Background:

• Multi-graph learning presents challenges for complex objects represented as bags of graphs.
• Supervised learning is hindered by the scarcity of labeled data, with only positive bags provided.

Purpose of the Study:

• To develop a robust framework for positive and unlabeled multi-graph learning (puMGL).
• To enhance the accuracy of classifying previously unseen graph bags.

Main Methods:

• puMGL framework selects informative subgraphs to create a feature space.
• It assigns confidence weights to unlabeled bags, identifying reliable negative bags.
• A margin graph pool is created from positive and reliable negative bags for pattern discovery and classifier training.

Main Results:

• The framework iteratively refines subgraph patterns and bag weights through a closed-loop process.
• Experimental results show puMGL's superior performance in classifying real-world complex objects.
• The method effectively handles the challenge of limited positive bags and numerous unlabeled bags.

Conclusions:

• The proposed puMGL framework offers an effective solution for positive and unlabeled multi-graph learning.
• This approach advances graph classification for complex objects with limited labeled data.
• puMGL demonstrates significant potential for real-world applications requiring accurate graph bag classification.