Multi-target regression via target specific features

摘要

Multi-target regression (MTR) has recently attracted great interest by the research community due to its capability of learning multiple related regression tasks. Existing approaches to MTR learn prediction models based on an identical set of features, which may be suboptimal since different targets may possess specific characteristics of their own. In this paper, we propose a method MTR-TSF that deals with the MTR tasks by learning target specific features (TSF). Firstly, a hierarchical clustering algorithm is applied to the output space of training data to reveal the similarities among multiple targets. Then, a classification and regression tree boosting method (CART-boosting) is used to generate a dependent similarity matrix for each target. Finally, target specific features are learned by querying the corresponding dependent similarity matrix and conducting clustering analysis on the training data. MTR-TSF method leverages the pertinent and discriminative features of each target and the dependence among multiple target to improve the overall performance of MTR. Experimental results on 18 real-world datasets demonstrate that MTR-TSF can achieve competitive performance against representative state-of-the-art MTR methods.