Meta-learning Method for Authomatic Selection of Algorithms for Text Classification

The paper presents a meta-learning approach for the task of textual document classification as an automatic selection of the optimal algorithm for creation of classifiers. Proposed method is based on the modified MUDOF algorithm, where regression model for prediction of optimizing parameters was replaced by a classification approach, using the kNN algorithm. This approach enables to increase quality and efficiency of semantic annotation procedures supported by text mining classification within collaborative eLearning system KP-Lab, which was taken as a prototype application. Design and implementation of the proposed meta-learning method using the JBowl Java library is described in detail. Finally, experimental results achieved by the meta-learning algorithms as well as their comparison with traditional ways used for text classification are presented and discussed.