结合目标特定特征和目标相关性的多目标回归

多目标回归旨在使用一组共同的输入变量来预测多个连续变量，其现有方法可归类为问题转换法和算法适应法.它的主要挑战在于如何对输入与输出空间的复杂关系进行建模，以及如何有效利用目标间的相关性.然而，现有的问题转换法很少同时考虑到这两方面.基于此，本文构建了一种问题转换法同时应对这两大挑战，提出了一种结合目标特定特征和目标相关性的多目标回归方法（Multi-Target Regression via Specific Features and Inter-Target Correlations，TSF-TC）.TSF-TC通过对分箱后的样本进行聚类分析构建目标特定特征从而对输入与输出空间的复杂关系进行建模，通过有选择性地堆叠单目标预测值揭示目标间的相关性.本文使用TSF-TC在18个多目标回归数据集上与现有多目标回归方法进行了对比实验，实验结果充分表明了TSF-TC的优势.

Multi-Target Regression via Specific Features and Inter-Target Correlations

Multi-target regression (MTR) aimed to predict multiple continuous variables using a common set of input variables,whose existing methods could be classified as problem transformation methods (PTM) and algorithm adaptation methods (AAM).Its main challenges were how to model the complex relationship between input and output space,and how to effectively utilize the correlation between targets.However,the existing PTM rarely took both aspects into consideration.So,this paper constructs a problem transformation method named TSF-TC combining target-specific features and Inter-Target correlations.TSF-TC constructs specific features to per target by conducting clustering analysis on the samples after binning,and then reveals the correlation between targets by selectively stacking single target prediction.Comparative experiments with existing multi-target regression methods on 18 datasets fully demonstrate the advantages of TSF-TC.