Comparison of Tree-Based Methods for Multi-target Regression on Data Streams

Abstract

Single-target regression is a classical data mining task that is popular both in the batch and in the streaming setting. Multi-target regression is an extension of the single-target regression task, in which multiple continuous targets have to be predicted together. Recent studies in the batch setting have shown that global approaches, predicting all of the targets at once, tend to outperform local approaches, predicting each target separately. In this paper, we explore how different local and global tree-based approaches for multi-target regression compare in the streaming setting. Specifically, we apply a local method based on the FIMT-DD algorithm and propose a novel global method, named iSOUP-Tree-MTR. Furthermore, we present an experimental evaluation that is mainly oriented towards exploring the differences between the local and the global approach.