Learning local instance correlations for multi-target regression

Abstract

Multi-target regression (MTR) refers to learning multiple relevant regression tasks simultaneously. Although much progress has been made in multi-target regression, there are still two challenging issues, that is, how to model the underlying relationships between input features and output targets, and how to explore inter-target dependencies. In this study, an effective algorithm named LLIC is proposed; it learns local instance correlations to reveal the relationships between features and output targets, and inter-target dependencies. First, an eminent instance selection method is adapted to directly work with multi-target data, constructing a collection of local instances for each instance. Then, in order to exploit the relationships between input features and output targets, and reveal inter-target dependencies, the collection of local instances is divided into two spaces, that is, a feature space and a target space. Implicit features of input features and targets are obtained in a statistical way. Finally, a final prediction model for each output target is trained on an expanded input space where the implicit features are treated as additional input variables. Extensive experiments on 18 benchmark datasets demonstrate that our proposed LLIC method can achieve competitive performance against representative state-of-the-art multi-target regression methods.