On Fusing Multiple Instance Selection Results

Instance selection is an important preprocessing step in data mining. Its aim is to filter out unrepresentative data samples from a given training dataset, which allow the classifier to perform better than the one without instance selection. Since various instance selection algorithms have been proposed in the literature, no study considers applying the information fusion principle to combine multiple instance selection results. This paper uses three well-known instance selection algorithms, which are IB3, DROP3, and GA, and their selection results are combined via the union, intersection, and multi-intersection strategies. Our experimental results based on 50 various domains of datasets show that the union between GA and DROP3 performs the best. However, it does not produce the largest reduction rate. On the other hand, if both classification accuracy and reduction rate are considered, the union between DROP3 and IB3 is the better choice.