Comparing Meta-learning Algorithms

Fabrício EnembreckBráulio Coelho Ávila

In this paper we compare the performance of KNOMA (Knowledge Mining Approach), a meta-learning approach for integration of rule-based classifiers, based on different rule inducers. Meta-learning approaches use a core learning algorithm for the generation of base classifiers that are further combined into a global one. This approach improves performance and scalability of data mining processes on large datasets. In a previous work we presented KNOMA, a meta-learning approach whose performance was evaluated using RIPPER as its core learning algorithm. Experiments have shown that the performance of KNOMA is comparable to that achieved with Bagging and Boosting. However, meta-learning is generally only sensitive to core algorithms used in the generation of base classifiers. KNOMA is a generic approach and can handle different rule-based inducers, although its advantages, drawbacks and use cases need to be precisely identified. We studied the variation of performance in the approach with base classifiers generated by two rule inducers (C45Rules and RIPPER) and also by C4.5. Interesting behaviors have been noticed in the experiments.

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: