Variable Selection in Bayesian Multiple Instance Regression using Shotgun Stochastic Search.

Area of Science:

• Machine Learning
• Statistical Modeling
• Bioinformatics

Background:

• Multiple Instance Learning (MIL) lacks interpretable models.
• Existing MIL methods often overlook model transparency and uncertainty quantification.

Purpose of the Study:

• To develop an interpretable Bayesian regression model for MIL.
• To simultaneously address instance and variable selection problems.
• To provide robust uncertainty quantification for MIL predictions.

Main Methods:

• A two-level hierarchical Bayesian regression model is proposed.
• A modified shotgun stochastic search algorithm is used for joint discrete space exploration.
• Instance and variable selection are integrated into the model framework.

Main Results:

• The model achieves high performance in variable and instance selection (AUC > 0.86).
• It demonstrates superior performance on the musk dataset for predicting molecular binding strengths.
• The method successfully identifies relevant variables for response modeling.

Conclusions:

• The proposed Bayesian MIL model offers enhanced interpretability and predictive accuracy.
• It effectively identifies key instances and variables, addressing critical needs in modern MIL.
• The approach provides a rigorous framework for uncertainty quantification in MIL.