LEARNING PARSIMONIOUS ENSEMBLES FOR UNBALANCED COMPUTATIONAL GENOMICS PROBLEMS.

Area of Science:

• Biomedical Sciences
• Computational Genomics
• Machine Learning

Background:

• Biomedical prediction is challenging due to incomplete knowledge and lack of consensus on ideal predictors.
• Ensemble methods improve prediction by combining base predictors, but parsimonious ensembles are valuable for interpretability.
• Existing ensemble selection algorithms like CES have parameter tuning challenges.

Purpose of the Study:

• To develop a novel heterogeneous ensemble selection approach using reinforcement learning (RL).
• To address the limitations of existing methods, particularly CES, in parameter selection and performance.
• To create parsimonious and accurate predictive models for biomedical problems.

Main Methods:

• Developed three RL-based strategies for constructing heterogeneous ensembles.
• Applied the RL ensemble selection approach to two unbalanced computational genomics problems: protein function prediction and splice site prediction.
• Compared the performance and parsimony of RL-based ensembles against the full set of base predictors and the CES algorithm.

Main Results:

• RL-based ensembles achieved substantial parsimony compared to the full predictor set while maintaining high classification power.
• The proposed RL method demonstrated a better balance of parsimony and predictive performance than the CES algorithm.
• Effectiveness was particularly noted on larger datasets, indicating scalability.

Conclusions:

• Reinforcement learning offers a systematic and mathematically sound methodology for ensemble selection in biomedical sciences.
• The developed RL strategies provide a robust alternative for constructing parsimonious and accurate predictive ensembles.
• This approach enhances both the predictive accuracy and interpretability of models in computational genomics.