Meta-aprendizagem aplicada a classificação de texto multirrótulo

Abstract

Classification is a predictive task of Data Mining that uses a set of data (instances), previously labeled, to train an algorithm that has the function of learning from the data presented and being able to predict the labels of new instances. The multi-label classification, in turn, differs from the traditional single-label classification in that it allows each instance of the dataset to be associ- ated with more than one label. Thus, in the domain of texts, characterized by diversity, volume, and increasing production, the multi-label classification plays an important role in allowing the maximum amount of information embedded in this data to be automatically extracted. Through texts, it is possible to identify the contents of great value for decision making, such as interests, opinions, and feelings. Some areas that explore ways to work with textual data are Natural Language Processing, Data Mining, and Machine Learning. The multi-label classification has a significant number of learning techniques available for its execution. However, finding the one that is most appropriate for a given dataset is not a trivial task, as it requires knowledge of techniques, consecutive experiments and, consequently, time. In this context, metalearning shows the relevance of its application when investigating ways to automate the process of se- lecting the best techniques for a given problem. Therefore, the objective of this work is to apply metalearning in the development of a method for the classification of multi-label texts, which seeks to select the best classification algorithm for each instance of the presented dataset. The experimental results demonstrated the effectiveness of the proposed method.