Loss Decomposition and Centroid Estimation for Positive and Unlabeled Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Positive and Unlabeled (PU) learning is a subfield of machine learning focused on binary classification tasks.
• Traditional supervised learning requires labeled positive and negative data, which is not always available.
• PU learning addresses scenarios where only positive and unlabeled data are accessible for training.

Purpose of the Study:

• To develop a robust PU learning algorithm that effectively handles the absence of negative training data.
• To propose a novel approach that converts PU learning into a risk minimization problem with one-sided label noise.
• To introduce methods that mitigate the impact of noisy labels inherent in PU learning settings.

Main Methods:

• Proposed a novel PU learning algorithm named "Loss Decomposition and Centroid Estimation" (LDCE).
• Decomposed the loss function of corrupted negative examples to isolate the impact of noisy labels.
• Introduced "Kernelized LDCE" (KLDCE) using the kernel trick, solvable via Alternative Convex Search (ACS) and Sequential Minimal Optimization (SMO).

Main Results:

• LDCE and KLDCE demonstrated effective handling of one-sided label noise in PU learning.
• Theoretical generalization error bounds were derived, showing convergence of generalization risk to empirical risk.
• Experimental results on synthetic, UCI benchmark, and real-world datasets confirmed top-level performance compared to existing methods.

Conclusions:

• LDCE and KLDCE are effective and robust algorithms for PU learning tasks.
• The proposed methods offer a significant advancement in handling datasets with limited or no negative labels.
• The findings suggest practical applicability in various real-world classification problems where negative data is scarce.