A composite splitting criterion using random sampling

The ever growing presence of data lead to a large number of proposed algorithms for classification and especially decision trees over the last few years. However, learning decision trees from large irrelevant datasets is quite different from learning small and moderate sized datasets. In practice, use of only small and moderate sized datasets is rare. Unfortunately, the most popular heuristic function gain ratio has a serious disadvantage towards dealing with large and irrelevant datasets. To tackle these issues, we design a new composite splitting criterion with random sampling approach. Our random sampling method depends on small random subset of attributes and it is computationally cheap to act on such a set in a reasonable time. The empirical and theoretical properties are validated by using 40 UCI datasets. The experimental result supports the efficacy of the proposed method in terms of tree size and accuracy.