Hybrid frameworks for classification : leveraging meta-learning strategies and ensemble approaches

Abstract

Standard classification tasks can face significant challenges when the model complexity is not well-matched to the problem. For instance, insufficient training data can result in high variance on the residuals, whereas using a model with an overly small set of features can lead to high bias. On the other extreme, tasks that involve excessively large training datasets or high-dimensional feature spaces can consume undesirable amounts of time and resources. Ensemble learning presents a compromising solution, training multiple relatively simple classifiers and then combining their outputs to achieve better classification performance, bringing a framework for accelerating the training procedure when the datasets involved are large. In this thesis, we discuss and explore several techniques for constructing classifiers within the ensemble learning framework, with a focus on multi-class classification tasks. We apply these methods to the task of digit recognition, evaluating their effectiveness on benchmark datasets. Throughout the experiments, we demonstrate the power of ensemble learning to deliver high-performing and computationally efficient solutions.

Date of Award	2024
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Ying Ju CHEN (Supervisor)