Improving the Predictive Performance of Online Decision Trees

Isvani FríasAlberto VerdeciaAndré Carlos Ponce de Leon Ferreira de Carvalho

Decision trees are powerful classification models for mining datastreams due to their characteristics, which include high interpretability, inducedby non-parametric techniques, fast and simple construction and high predictiveaccuracy. However, in many current data mining applications new data arecontinuously arriving in streams, frequently making previously induced modelsoutdated. This paper introduces a new prediction technique for decision treeinduction, suitable for learning from high-speed data streams. The new techniquecombines the predictions from Na¨?ve Bayes and majority class by usingensemble techniques. The empirical study shows that, when used in the leafnodes of the Very Fast Decision Tree technique, the new technique is able tosignificantly increase predictive accuracy when compared with other popularprediction techniques for data streams, being also able to process training instanceswith limited computational resources.

Caso o link acima esteja inválido, faça uma busca pelo texto completo na Web: Buscar na Web

Biblioteca Digital Brasileira de Computação - Contato:
     Mantida por: