Soft Monotonic Constraint Support Vector Regression

This paper proposes a model for learning soft-monotonic regression functions in the presence of imperfect domain knowledge. It proposes an extension to support vector regression (SVR) wherein a new hardness parameter is introduced to configure the degree of monotonicity. The model supports multiple monotonicity constraints over multiple input dimensions simultaneously. The proposed model has been validated on synthetic datasets as well as on benchmark datasets obtained from real world problems. The results show that our model has better extrapolation capabilities than SVR. The results also demonstrate the ability of the model to generalize over multiple input dimensions.