Interpretable Ensembles of Classifiers for Uncertain Data With Bioinformatics Applications.

Area of Science:

• Computational biology
• Machine learning
• Bioinformatics

Background:

• Data uncertainty is a significant challenge in classification tasks.
• Existing ensemble methods for uncertain categorical features show promise.
• Effective handling of data uncertainty is crucial for accurate predictions in biological applications.

Purpose of the Study:

• To develop novel ensemble approaches for handling data uncertainty in classification.
• To improve predictive performance in datasets with uncertain features.
• To introduce new methods for interpreting ensemble models.

Main Methods:

• Proposed two new ensemble strategies: biased instance selection and biased feature sampling for Random Forest node splitting.
• Applied these methods to classify ageing-related genes and predict drug side effects using uncertain interaction data.
• Developed new interpretation techniques for ensembles of Naive Bayes classifiers.

Main Results:

• Ensembles utilizing the proposed biased sampling approaches demonstrated superior predictive performance.
• Significant improvements were observed, particularly in Random Forest models employing the most advanced uncertainty handling techniques.
• New interpretation methods provided insights into ensemble predictions for biological datasets.

Conclusions:

• The developed ensemble methods effectively address data uncertainty in classification.
• Biased sampling strategies enhance the accuracy of machine learning models for biological data.
• The proposed interpretation techniques offer valuable tools for understanding complex biological predictions.