Partially linear support vector machines applied to the prediction of mine slope movements

We propose a partially linear version of the SVMs, PL-SVM, which uses a kernel composed of a linear kernel in which a number of variables participate, and a nonlinear kernel which affects the other variables. This approach enables a linear component in a subset of variables to be modeled. The resulting models are true SVMs and so existing learning algorithms can be used. This approach can be applied to other kernel methods such as kernel discriminant analysis, kernel principal components analysis, etc. We used an autoregressive PL-SVM with a view to predicting monthly movement in a mine slope with an impact on the safety of the mining operation. In our problem, the PL-SVM improves on the results of other autoregressive approaches, including those for the classical non-parametric partially linear models.