Degree Type


Date of Award


Degree Name

Doctor of Philosophy


Industrial and Manufacturing Systems Engineering


Industrial and Manufacturing Systems Engineering

First Advisor

Sigurdur Olafsson


Ensemble learning is a popular classification method where many individual simple learners contribute to a final prediction. Constructing an ensemble of learners has been shown to consistently improve prediction accuracy over a single learner. The most common types of ensembles include: bootstrap aggregated (bagged), boosted, and stacked. Each are different, yet has the same foundation of combining multiple learners.

In this dissertation, we focus our attention to bagged ensembles; namely we propose a generalization by way of model weighting. The new method is motivated by the potential instability of averaging predictions of trees that may be of highly variable quality. To alleviate this, we replace the usual arithmetic average with a Cesaro average for weighted trees in the random forest. We provide both a theoretical analysis that gives exact conditions under which we would expect this weighted ensemble approach to do well, and numerical analysis that shows the new approach is competitive to other bagged ensembles when training a classification model on numerous realistic data sets.

Going a step further we generalize our weights such that we allow simultaneous control over bias and variance. In particular, we introduce a regularization term that controls the variance reduction for bagged ensembles. Therefore, a new tunable weighted bagged ensemble framework is proposed, resulting in a very flexible method for classification. Using this methodology, we explore the impact tunable weighting has on the votes of each learner in an ensemble.

To aid in the applicability of this body of work, the author discusses an R package that allows users to implement our proposed weighting scheme to arbitrary bagged ensembles. The package provides tools for constructing tunable bagged ensembles in the form of weights and is titled wbensembleR.

Copyright Owner

Hieu Trung Pham



File Format


File Size

86 pages