Using Turning Point Detection to Obtain Better Regression Trees

Abstract

The issue of detecting optimal split points for linear regression trees is examined. A novel approach called Turning Point Regression Tree Induction (TPRTI) is proposed which uses turning points to identify the best split points. When this approach is used, first, a general trend is derived from the original dataset by dividing the dataset into subsets using a sliding window approach and a centroid for each subset is computed. Second, using those centroids, a set of turning points is identified, indicating points in the input space in which the regression function, associated with neighboring subsets, changes direction. Third, the turning points are then used as input to a novel linear regression tree induction algorithm as potential split points. TPRTI is compared in a set of experiments using artificial and real world data sets with state-of-the-art regression tree approaches, such as M5. The experimental results indicate that TPRTI has a high predictive accuracy and induces less complex trees than competing approaches, while still being scalable to cope with larger datasets.