Um método evolucionário para classificação hierárquica

Abstract

The traditional (flat) classification task has been extensively explored in several areas of knowledge. In the last decade, however, research efforts turn to the task of hierarchical classification, a specialization of the previous task. The main difference between the two is that in hierarchical classification examples are associated with classes organized into a predefined class hierarchy, whereas in flat classification no class order is enforced. Currently, two main approaches are used to generate hierarchical classification models: (i) the local approach, also known as top-down, and (ii) the global approach, also called big-bang. The first strategy uses flat classifiers to generate models for each node or parent node of the hierarchy, and take into account local information of the hierarchy when developing these models, ending with a set of models. For each new example, these models are used to predict their most generic classes, and then the most specific ones. The global approach, in contrast, typically modifies a flat classifier to add information about the hierarchy for building the classification model. Many traditional classification algorithms are being successfully adapted for hierarchical classification, such as SVM and Naive Bayes, while new algorithms proposals are yet underexplored. In this direction, this work proposes HCGA (Hierarchical Classification Genetic Algorithm), an evolutionary method for hierarchical classification whose main contribution is to merge both local and global information of the hierarchy, using a top-down approach for both creating the class models (training) and classifying new examples (test). This represents a major innovation, since the current top-down methods use only local information in training, and are not aware of the hierarchy structure when building models. Hierarchical problems have been exploited mainly in the areas of text classification and bioinformatics. Hence, the proposed method was applied to four datasets regarding protein function prediction and a well-known text classification dataset. The results obtained were compared to five other algorithms commonly used in hierarchical classification, obtaining promising results specially in the deeper levels of the hierarchy.