AdaSampling for Positive-Unlabeled and Label Noise Learning With Bioinformatics Applications

Area of Science:

• Machine Learning
• Bioinformatics
• Computational Biology

Background:

• Supervised learning requires accurate class labels, which are often missing or corrupted in real-world applications.
• Positive-Unlabeled (PU) learning addresses scenarios with only labeled positive instances and unlabeled data.
• Learning with Label Noise (LN) is crucial when training data contains mislabeled instances.

Purpose of the Study:

• To introduce Adaptive Sampling (AdaSampling), a novel framework for both PU learning and learning with class label noise.
• To develop a method that iteratively estimates mislabeling probability and reduces reliance on mislabeled data.
• To construct highly generalizable models even with a high proportion of noisy labels.

Main Methods:

• AdaSampling employs an iterative adaptive sampling procedure to estimate class mislabeling probabilities.
• The framework progressively refines model training by down-weighting or excluding potentially mislabeled instances.
• The method's performance is validated using simulations, benchmark datasets, and novel bioinformatics applications.

Main Results:

• AdaSampling demonstrates effectiveness in both PU learning and learning with label noise scenarios.
• The proposed method successfully constructs generalizable models despite substantial label noise.
• Comparative analyses show AdaSampling outperforms common alternative approaches.

Conclusions:

• AdaSampling offers a robust solution for machine learning tasks with incomplete or noisy labels.
• The framework shows significant promise in bioinformatics, specifically for identifying kinase-substrates and predicting transcription factor targets.
• AdaSampling facilitates the development of more reliable predictive models in data-scarce or noisy biological datasets.