Relabeling algorithm for retrieval of noisy instances and improving prediction quality.

Area of Science:

• Machine Learning
• Data Mining
• Predictive Analytics

Background:

• Noisy data instances with binary outcomes can significantly degrade prediction quality in machine learning models.
• Existing methods often prioritize classification accuracy, potentially overlooking crucial aspects like knowledge generalization and data confidence.

Purpose of the Study:

• To present a novel relabeling algorithm designed for the retrieval and correction of noisy instances in datasets with binary outcomes.
• To enhance prediction quality by focusing on knowledge generalization and confidence, rather than solely on classification accuracy.

Main Methods:

• The algorithm iteratively retrieves, selects, and relabels data instances, effectively transforming the decision space.
• A confidence index was developed, integrating classification accuracy, prediction error, dataset impurities, and cluster purities.
• The approach was validated on benchmark UCI datasets and bladder cancer immunotherapy data.

Main Results:

• A subset of stable instances (7%-51%) with high confidence (64%-99.44%) was identified, alongside the most noisy instances.
• The relabeled instances and their confidence indexes were validated by domain experts.
• The algorithm demonstrated successful application on diverse datasets, including medical and benchmark data.

Conclusions:

• The proposed relabeling algorithm effectively improves prediction quality by addressing noisy instances.
• The emphasis on knowledge generalization and confidence provides a robust measure of data reliability.
• The algorithm shows potential for broad applicability across medical, industrial, and service domains with modifications.