Combining decisions of multiple rules

Several existing combination rules are extended in many ways and tested on sets of rules generated by different learning algorithms in several classification problems, indicating that the selection of the appropriate combination rule depends on the classification problem as well as on the learning algorithm.

Abstract

The idea of learning of multiple knowledge is to generate several theories instead of one theory for the classification of new objects, hoping that combining of answers of more theories will result in better performance. However, different authors use different strategies for combining multiple rules. In this paper several existing combination rules are extended in many ways and tested on sets of rules generated by different learning algorithms in several classification problems. Results indicate that the selection of the appropriate combination rule depends on the classification problem as well as on the learning algorithm. However, the naive combination of probabilities seems to be superior to other existing combination rules.