Instance elimination strategy for non-convex multiple-instance learning using sparse positive bags.

Area of Science:

• Machine Learning
• Computer Science
• Data Science

Background:

• Multiple Instance Learning (MIL) often faces challenges with imbalanced data, particularly when positive bags are sparse.
• Existing methods may struggle with the computational demands of large-scale MIL problems.

Purpose of the Study:

• To propose a novel MIL method, Small Sphere and Large Margin MIL (SSLM-MIL), to address data imbalance in sparse positive bag scenarios.
• To enhance the computational efficiency of solving non-convex optimization problems within MIL frameworks.

Main Methods:

• Developed the SSLM-MIL method incorporating a large margin to ensure at least one positive instance per positive bag.
• Utilized the concave-convex procedure (CCCP) for non-convex optimization.
• Introduced novel safe instance screening rules for both inner solvers and inter-iteration propagation within CCCP.

Main Results:

• The proposed SSLM-MIL method effectively handles sparse positive bags and data imbalance.
• The integrated screening rules significantly reduce the computational scale and time for CCCP.
• Demonstrated the safety and effectiveness of the approach across thirty-one benchmark datasets.

Conclusions:

• SSLM-MIL offers an effective solution for imbalanced MIL problems with sparse positive instances.
• The novel screening strategy enhances the practicality of using CCCP for non-convex MIL.
• This work represents the first application of safe instance screening to non-convex hypersphere support vector machines.