Multi-relational Naive Bayesian classifier based on mutual information

To improve the accuracy of multi-relational Naive Bayesian classifiers, the existing pruning methods were discussed and the attribute filter criterion was upgraded based on mutual information to deal with multi-relational data directly. On the basis of the tuple ID propagation method and counting methods towards tuple, the filter method based on extended mutual information was given, and a multi-relational Naive Bayesian classifier based on mutual information (MI-MRNBC) was implemented. Experimental results show that, in a multi-relational domain, with the help of the attribute filter based on extended mutual information, the classifier can give a better accuracy without the increase of time complexity. In extraordinary instances, the multi-relational classification degenerates into a single relational one, which extremely decreases the cost of classification.