View selection in multi-view stacking: choosing the meta-learner

Area of Science:

• Machine Learning
• Bioinformatics
• Statistical Modeling

Background:

• Multi-view stacking integrates information from diverse feature sets (views) for object analysis.
• Previous work demonstrated stacked penalized logistic regression's utility in identifying predictive data views.
• This study extends multi-view stacking research by exploring various meta-learner algorithms.

Purpose of the Study:

• To evaluate the view selection and classification performance of seven different meta-learner algorithms within the multi-view stacking framework.
• To identify optimal meta-learners for applications requiring both accurate view selection and high classification performance, particularly in gene expression data analysis.

Main Methods:

• Implemented a multi-view stacking framework using seven distinct meta-learner algorithms.
• Conducted simulations and analyzed two real-world gene-expression datasets to assess performance.
• Evaluated algorithms based on their ability to perform view selection and improve classification accuracy.

Main Results:

• Nonnegative lasso, nonnegative adaptive lasso, and nonnegative elastic net demonstrated superior performance for both view selection and classification accuracy.
• The choice among these three top-performing meta-learners depends on specific research requirements.
• Other evaluated meta-learners (nonnegative ridge regression, forward selection, stability selection, interpolating predictor) offered limited advantages.

Conclusions:

• For research prioritizing both view selection and classification accuracy, nonnegative lasso, adaptive lasso, and elastic net are recommended meta-learners in multi-view stacking.
• The selection of the best meta-learner among these three is context-dependent.
• The study provides valuable insights for optimizing multi-view stacking in bioinformatics and related fields.