Learning With Label Proportions via NPSVM.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Learning from label proportions (LLP) is an emerging area in machine learning.
• Existing LLP methods often rely on transductive learning, leading to complex optimization problems.

Purpose of the Study:

• To propose a novel algorithm, LLPs via nonparallel support vector machine (LLP-NPSVM), to address the challenges in learning from label proportions.
• To develop an efficient and accurate method for instance-level prediction using only bag-level label proportions.

Main Methods:

• The proposed LLP-NPSVM determines instance labels using two nonparallel hyper-planes, supervised by label proportion information.
• This approach avoids the transductive learning framework and instead uses a large margin clustering concept.
• The algorithm is efficiently solved using two fast sequential minimal optimization paths iteratively.

Main Results:

• The LLP-NPSVM procedure demonstrates finite termination and monotonic decrease, ensuring stability.
• Experimental results show rapid convergence and robust numerical stability.
• The algorithm achieves superior accuracy compared to several recent methods in most cases.

Conclusions:

• LLP-NPSVM offers an effective and efficient solution for learning from label proportions.
• The method provides a competitive alternative to existing LLP techniques, particularly in scenarios with weak labels.