Kernekoncepter
The choice of meta-learner in multi-view stacking can significantly impact the trade-off between classification accuracy and view selection performance.
Resumé
The article investigates the performance of seven different meta-learners in the context of multi-view stacking (MVS), a framework for combining information from different feature sets (views) to build accurate classification models. The authors compare the meta-learners in terms of classification accuracy, true positive rate (TPR), false positive rate (FPR), and false discovery rate (FDR) in view selection, using both simulations and two real gene expression data sets.
The key findings are:
The nonnegative lasso, adaptive lasso, elastic net, and nonnegative forward selection (NNFS) generally showed comparable classification performance, while nonnegative ridge regression, stability selection, and the interpolating predictor performed noticeably worse in some conditions.
Among the well-performing meta-learners, model sparsity was associated with a lower FPR but also a lower TPR in view selection. However, there were situations where the sparser meta-learners obtained both a low FPR and a high TPR, particularly when the features from different views were uncorrelated.
Even when the FPR was very low, the FDR was often high, especially with a small sample size (n=200).
In the real data applications, the nonnegative lasso, adaptive lasso, elastic net, and NNFS selected a similar number of views, but the stability of the selected views varied, with the nonnegative lasso and adaptive lasso being the most stable.
The authors conclude that if both view selection and classification accuracy are important, the nonnegative lasso, adaptive lasso, and elastic net are suitable meta-learners, with the choice depending on the specific research context.
Statistik
"The sample size is either n=200 or n=2000."
"There are either V=30 or V=300 views, with each view containing either mv=250 or mv=2500 features."
"The population correlation between features from the same view is ρw=0.1, 0.5, or 0.9, and the population correlation between features from different views is ρb=0, 0.4, or 0.8."
Citater
"If both view selection and classification accuracy are important to the research at hand, then the nonnegative lasso, nonnegative adaptive lasso and nonnegative elastic net are suitable meta-learners."
"Exactly which among these three is to be preferred depends on the research context."
"The remaining four meta-learners, namely nonnegative ridge regression, nonnegative forward selection, stability selection and the interpolating predictor, show little advantages in order to be preferred over the other three."