Stacking Model-Based Classifiers for Dealing With Multiple Sets of Noisy Labels.

Area of Science:

• Machine Learning
• Healthcare Analytics
• Data Science

Background:

• Supervised learning with multiple noisy labels is a significant challenge in healthcare analytics.
• Discrepancies in labels arise from multiple annotators with varying expertise and subjectivity.
• Addressing noisy labels is crucial for accurate classification in medical data.

Purpose of the Study:

• To develop a novel ensemble methodology for classification tasks with multiple sets of noisy labels.
• To improve predictive performance in healthcare analytics by effectively handling label discrepancies.
• To automatically infer annotator expertise levels as a byproduct of the learning process.

Main Methods:

• An ensemble methodology combining model-based classifiers trained on individual noisy label sets.
• Eigenvalue Decomposition Discriminant Analysis (EDDA) used for base learner definition.
• Six distinct averaging strategies proposed for combining base learners, including data-driven and information-dependent approaches.

Main Results:

• The proposed ensemble methodology demonstrates improved predictive performance compared to existing methods.
• A simulation study and real-world data application validate the effectiveness of the approach.
• The method successfully infers annotator expertise levels without prior knowledge.

Conclusions:

• The novel ensemble approach effectively addresses supervised learning challenges posed by multiple noisy labels in healthcare.
• The methodology offers a robust solution for improving classification accuracy in medical data analysis.
• The ability to infer annotator expertise adds valuable insight into data quality and reliability.