A Classifier Selection Strategy for Lazy Bayesian Rules Based on Local Accuracy Estimation

Lazy Bayesian rule (LBR) is a novel classification method of high predictability. However, its classifier selection strategy is somewhat simple, in that it always uses the most specific one to make the final decision. In this paper, we suggest to use the one with the highest estimated local accuracy at the target instance point instead. To materialize this idea, this paper proposes an efficient mechanism for estimating the local accuracy of a local classifier. This mechanism can be easily integrated into the LBR algorithm, and therefore leads to a classifier selection strategy for LBR. Experimental results have shown that this classifier selection strategy can reduce the error rate of the original LBR algorithm averagely over a variety of domains.