Family Rank: a graphical domain knowledge informed feature ranking algorithm.

Area of Science:

• Computational Biology
• Bioinformatics
• Statistical Modeling

Background:

• Feature selection methods can overfit training data in prediction models with many features and small sample sizes, leading to the selection of irrelevant features.
• Incorporating domain knowledge into feature selection is a potential strategy to mitigate overfitting.
• The Family Rank algorithm is introduced, which ranks features using both graphical domain knowledge and empirical data scores.

Purpose of the Study:

• To introduce and evaluate the Family Rank algorithm for feature selection.
• To demonstrate the algorithm's ability to mitigate overfitting in prediction models.
• To assess the performance of Family Rank compared to existing graph-based ranking algorithms.

Main Methods:

• The Family Rank algorithm combines graphical domain knowledge with feature scores derived from empirical data.
• A simulated dataset was used to compare Family Rank against state-of-the-art graph-based ranking algorithms.
• A real-world application in oncology was explored to demonstrate practical utility.

Main Results:

• Family Rank outperformed other graph-based ranking algorithms on a simulated dataset.
• The algorithm reduced the required sample size for detecting true predictors by 2- to 3-fold.
• The study explored a real-world application of Family Rank in oncology.

Conclusions:

• Family Rank offers an effective approach to feature selection by integrating domain knowledge, potentially improving prediction model performance.
• The algorithm demonstrates superior performance in scenarios with limited sample sizes.
• Family Rank provides a valuable tool for researchers, with an implementation available for use.