Relaxing instance boundaries for the search of splitting points of numerical attributes in classification trees

We propose a simple heuristic partition method (HPM) of classification tree to improve efficiency in the search for splitting points of numerical attributes. The proposal is motivated by the idea that the selection process of candidates in the splitting point selection can be made more flexible as to achieve a faster computation while retaining classification accuracy. We compare the performance of the HPM against Fayyad’s method, as the latter is the improved version of the standard C4.5 algorithm on the search of splitting points. We demonstrate that HPM is more efficient, in some cases by as much as 50%, while producing essentially the same classification for six different data sets. Our result supports the relaxation of instance boundaries (RIB) as a valid approach that can be explored to achieve more efficient computations.