The design and implementation of a method for opinion classification based on machine learning algorithms

Opinion classification, which utilizes methods, techniques and tools from natural language processing, machine learning and statistics, is a research field to determine subjective information in the text documents. In this thesis, opinion mining is regarded as a text classification problem to build an efficient opinion classification scheme based on machine learning methods. The developed methods can be grouped into three categories. First, a feature selection method based on genetic rank aggregation, which integrates individual filter-based feature selection methods in an effective way, is presented to overcome the high dimensionality and sparsity problems encountered in text classification. Secondly, a classifier ensemble combination rule based on multi-objective differential evolution algorithm is presented so that the base learning algorithms of the classifier ensemble can contribute to the final outcome of the ensemble according to their predictive performance. Thirdly, an ensemble pruning scheme is presented to obtain an appropriate subset of classifiers from the ensemble. In the proposed pruning scheme, consensus clustering and metaheuristic search are utilized. A novel classification architecture is proposed based on the developed methods. With this framework, developed methods are combined in an effective way and the predictive performance of the framework is enhanced compared to the individual performances of the standard and the developed methods.