Learning Predictive Interactions Using Information Gain and Bayesian Network Scoring.

Area of Science:

• Statistics and Machine Learning
• Genetics and Genomics
• Bioinformatics

Background:

• High-dimensional data, common in fields like Genome-wide Association Studies (GWAS), presents challenges in identifying predictive variables with small marginal effects.
• Existing methods struggle to distinguish true interactions from variables with strong individual effects.
• Epistatic interactions between single nucleotide polymorphisms (SNPs) are crucial for understanding disease status in GWAS.

Purpose of the Study:

• To develop a method that accurately identifies interacting variables in high-dimensional datasets.
• To differentiate true interactions from variables with strong marginal effects.
• To improve the discovery of predictive relationships in complex datasets.

Main Methods:

• The MBS-IGain strategy combines information gain and Bayesian network scoring.
• Candidate interactions are identified by assessing if variables provide more information together than separately.
• Bayesian network scoring is used to validate the likelihood of candidate interactions as models.

Main Results:

• MBS-IGain significantly outperformed nine previous methods in locating and exactly identifying interacting predictors on simulated datasets.
• The method yielded new and substantiated findings when applied to a real GWAS Alzheimer's dataset.
• MBS-IGain demonstrated high effectiveness in finding interactions within high-dimensional data.

Conclusions:

• MBS-IGain is a highly effective tool for identifying interactions in high-dimensional datasets.
• Accurate interaction identification is crucial for learning causes and performing prediction/classification with abundant high-dimensional data.
• The method offers a significant advancement for analyzing complex biological and statistical data.