A Boosting Approach to Exploit Instance Correlations for Multi-Instance Classification.

Area of Science:

• Machine Learning
• Computer Science
• Data Mining

Background:

• Multi-instance (MI) classification presents challenges in learning from data where labels are assigned to sets of instances.
• Existing methods often rely on probabilistic assumptions or complex feature fusion strategies.

Purpose of the Study:

• To develop a flexible and concise Boosting model for MI classification.
• To improve MI classification performance by directly localizing key instances and fusing their scores.

Main Methods:

• Integration of L_p-norm for localizing 'witness' instances within multi-instance data.
• Development of a Boosting framework using exponential loss optimization and direct score fusion from instance features.
• Application of gradient and Newton descent optimizations for deriving weak learners, incorporating instance correlations.

Main Results:

• The proposed L_p-norm-localized Boosting approach significantly enhances MI classification performance.
• Achieved the highest MI classification accuracy on 7 out of 10 benchmark datasets compared to state-of-the-art methods.
• Demonstrated effective fusion of instance-level information for robust bag-level classification.

Conclusions:

• The L_p-norm-localized Boosting method provides a powerful and efficient solution for multi-instance classification tasks.
• The approach offers a flexible alternative to probabilistic models, directly leveraging instance features.
• Experimental results validate the superiority of the proposed method in improving classification accuracy.