Adjusted geometric-mean: a novel performance measure for imbalanced bioinformatics datasets learning.

Area of Science:

• Bioinformatics
• Machine Learning
• Computational Biology

Background:

• Imbalanced datasets are prevalent in bioinformatics tasks like gene prediction and drug discovery.
• This imbalance often results in models with high specificity but low sensitivity, hindering accurate positive class identification.

Purpose of the Study:

• To address the limitations of existing performance measures in class imbalance learning for bioinformatics.
• To introduce a novel metric, Adjusted Geometric-mean (AGm), that optimizes the trade-off between sensitivity and specificity.

Main Methods:

• Evaluation of existing class imbalance learning methods and their performance metrics.
• Introduction and application of the Adjusted Geometric-mean (AGm) metric.
• Experimental validation on ten real-world imbalanced bioinformatics datasets.

Main Results:

• Existing metrics can yield suboptimal models even after applying class imbalance techniques.
• The AGm metric demonstrated a reduced loss in specificity compared to traditional metrics when enhancing sensitivity.
• AGm proved more effective in achieving high sensitivity with minimal specificity reduction.

Conclusions:

• The AGm metric is better suited for imbalanced bioinformatics datasets where maximizing sensitivity while preserving specificity is crucial.
• This new metric facilitates the development of more effective prediction models in challenging bioinformatics applications.
• AGm offers a refined approach to evaluating models in data-imbalanced scenarios.