Large-margin Distribution Machine-based regression

This paper presents an efficient and robust Large-margin Distribution Machine formulation for regression. The proposed model is termed as ‘Large-margin Distribution Machine-based Regression’ (LDMR) model, and it is in the spirit of Large-margin Distribution Machine (LDM) (Zhang and Zhou, in: Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, ACM, 2014) classification model. The LDM model optimizes the margin distribution instead of minimizing a single-point margin as is done in the traditional SVM. The optimization problem of the LDMR model has been mathematically derived from the optimization problem of the LDM model using an interesting result of Bi and Bennett (Neurocomputing 55(1):79–108, 2003). The resulting LDMR formulation attempts to minimize the ϵ-insensitive loss function and the quadratic loss function simultaneously. Further, the successive over-relaxation technique (Mangasarian and Musicant, IEEE Trans Neural Netw 10(5):1032−1037, 1999) has also been applied to speed up the training procedure of the proposed LDMR model. The experimental results on artificial datasets, UCI datasets and time-series financial datasets show that the proposed LDMR model owns better generalization ability than other existing models and is less sensitive to the presence of outliers. © 2018, Springer-Verlag London Ltd., part of Springer Nature.