Restructuring of Hoeffding Trees for Trapezoidal Data Streams

Abstract

Trapezoidal Data Streams are an emerging topic, where not only the data volume increases, but also the data dimension, i.e. new features emerge. In this paper, we address the challenges that arise from this problem by providing a novel approach to restructure and prune Hoeffding trees. We evaluate our approach on synthetic datasets, where we can show that the approach significantly improves the performance compared to the baseline of an adjusted Hoeffding tree algorithm without restructuring and pruning.