Improving the Accuracy of Incremental Decision Tree Learning Algorithm via Loss Function

Hoeffding´s bound (HB) has been widely used for node splitting in incremental decision tree algorithms. Many decision-tree algorithms adopt a sliding-window technique to detect concept drift when mining changing data streams. This paper presents a novel node-splitting approach that replaces the traditional HB with a new measure. The new measure is derived from a loss function applied in a cache-based classifier within a sliding window during incremental decision tree learning. Replacing the use of HB with this new bound is proposed for growing a Hoeffding decision tree that adapts to concept drifts detected in the data stream, thus improving the accuracy of prediction. The experimental results show that this new method has the potential to achieve better performance with fine tuning of the sliding window size.