[en] We present HEUQ, a novel heterogeneous ensemble with uncertainty quantification (UQ) for churn prediction.
It contributes to the extensive literature on ensemble methods and churn prediction with the incorporation of UQ, which is overlooked in extant studies, resulting in novel research questions. Our work aims at addressing these questions.
We decompose uncertainty into aleatoric and epistemic uncertainty, expressed as the average divergence between predictive probability distributions of the ensemble and its models.
Intuitively, a low epistemic uncertainty indicates consensus in the individual model's prediction, and serves as grounding for HEUQ's development.
We perform extensive experiments on a real-life dataset, including an ablation study to evaluate individual learners' importance in the ensemble. Results show that HEUQ achieves better predictive performance than other baselines, and is associated with
lowest uncertainty.
HEUQ's performance is also stable to transformations (PCA) and distortion (introduced by Random Gaussian projections), indicating its robustness.
% \textcolor{blue}{\textbf{Management applications}}.
We believe that our approach of uncertainty quantification could open up a new paradigm for classifier selection in an ensemble. In addition to its scientific contributions, our study has immediate practical relevance. We show how \texttt{HEUQ} can be used in a real-life management application.
