Statistical Learning Methods for Predicting Ordinal and Count Outcomes in Genomics Data

With advancements in statistical modeling and the growing availability of high-dimensional genomics data, researchers have developed numerous methods to address the challenges and opportunities presented in such data. While many classification or regression techniques exist, their performance often depends heavily on the evaluation criteria used, and it is rarely clear a priori which technique will perform best for a given application. To address this uncertainty, we consider ensemble approaches that leverage bagging and rank aggregation techniques, applied separately to classification and regression problems. In the first part of this work, we introduce an ensemble classifier designed to optimize classification of ordinal outcomes in high-dimensional data. This classifier integrates several existing ordinal classification algorithms, improving predictive performance across multiple evaluation metrics. The second part focuses on ensemble prediction for count outcomes in high-dimensional settings. Similar to the classification framework, this regression ensemble combines multiple existing algorithms to optimize predictive performance across various evaluation criteria. Through extensive simulation studies and real-world genomic applications, we demonstrate that our ensemble methods consistently outperform or closely match the best-performing individual algorithms in both classification and regression contexts. These findings highlight that, when addressing the complexities of high-dimensional data, adopting an ensemble approach that integrates multiple models may provide superior and more reliable performance than relying on a single algorithm.