In this thesis a number of hierarchical ensemble classification approaches are proposed as a solution to the multi-class classification problem. The central idea is that a more effective classification can be produced if a “coarse-grain” classification (directed at groups of classes) is first conducted followed by increasingly more “fine-grain” classifications. The Hierarchical ensemble classification model comprises a set of base classifiers held within the nodes of the hierarchy (one classifier per node). Nodes near the root hold classifiers designed to discriminate between groups of class labels while the leaves hold classifiers designed to distinguish between individual class labels. Two types of hierarchy (structures) are considered, Binary Tree (BT) hierarchies and Directed Acyclic Graph (DAG) hierarchies. With respect to the DAG structure, two alternative DAG structures to support the generation of the desired hierarchical ensemble classification model are considered: (i) rooted DAG, and (ii) non-rooted DAG. The main challenges are: (i) how best to distribute class labels between nodes within the hierarchy, (ii) how to address the “successive mis-classification” issue associated with hierarchical classification where if a mis-classication occurs early on in the process (near the root of the hierarchy) there is no possibility of rectifying this error later on in the process, and (iii) how best to determine the starting node within the non-rooted DAG approach. To address the first issue different techniques, based on the concepts of clustering, splitting, and combination, are proposed. To address the second and the third issues the idea is to utilise probability or confidence values associated with Naive Bayes and CARM classifiers respectively to dictate whether single or multiple paths should be followed at each hierarchy node, and to select the best starting DAG node with respect to the non-rooted DAG approach.
展开▼
