Improving the hierarchical classification of protein functions With swarm intelligence

This thesis investigates methods to improve the performance of hierarchical classification. In terms of this thesis hierarchical classification is a form of supervised learning, where the classes in a data set are arranged in a tree structure. As a base for our new methods we use the TDDC (top-down divide-and-conquer) approach for hierarchical classification, where each classifier is built only to discriminate between sibling classes.

Firstly, we propose a swarm intelligence technique which varies the types of classifiers used at each divide within the TDDC tree. Our technique, PSO/ACO-CS (Particle Swarm Optimisation/Ant Colony Optimisation Classifier Selection), finds combinations of classifiers to be used in the TDDC tree using the global search ability of PSO/ACO.

Secondly, we propose a technique that attempts to mitigate a major drawback of the TDDC approach. The drawback is that if at any point in the TDDC tree an example is misclassified it can never be correctly classified further down the TDDC tree. Our approach, PSO/ACO-RO (PSO/ACO-Recovery Optimisation) decides whether to redirect examples at a given classifier node using, again, the global search ability of PSO/ACO.

Thirdly, we propose an ensemble based technique, HEHRS (Hierarchical Ensembles of Hierarchical Rule Sets), which attempts to boost the accuracy at each classifier node in the TDDC tree by using information from classifiers (rule sets) in the rest of that tree. We use Particle Swarm Optimisation to weight the individual rules within each ensemble.

We evaluate these three new methods in hierarchical bioinformatics datasets that we have created for this research. These data sets represent the real world problem of protein function prediction.

We find through extensive experimentation that the three proposed methods improve upon the baseline TDDC method to varying degrees. Overall the HEHRS and PSO/ACO- CS-RO approaches are most effective, although they are associated with a higher computational cost.