A novel support vector regression algorithm incorporated with prior knowledge and error compensation for small datasets

To solve the modeling problem under the conditions where the measured data are insufficient but biased prior knowledge from a simulator is available, we propose a novel multi-scale ν -linear programming support vector regression ( ν -LPSVR) called ν -MPESVR. The proposed algorithm constructs a nested support vector regression model, which incorporates prior knowledge into ν -LPSVR, compensates for the errors between prior knowledge data and the measured data, and simultaneously achieves small training error for both the prediction model and the error compensation model. Considering that measured data may exist in multiple feature spaces, we extend the algorithm to multi-scale ν -LPSVR to achieve accurate modeling for complex problems. In addition, a strategy for parameter selection for ν -MPESVR is presented in this paper. The performance of the proposed algorithm is estimated in a synthetic example and a practical application. The performances of all models are evaluated with the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2). Taking the three groups of experiments in the synthetic example as an instance, we find that the ν -MPESVR performs better, and it can still maintain high accuracy when the biases of prior knowledge data change (RMSE values of 0.1962, 0.1904, and 0.2261, MAE values of 0.1396, 0.1375, and 0.1623, and R2 values of 0.9919, 0.9923, and 0.9892 for the three groups of experiments, respectively). The experimental results indicate that the proposed algorithm can obtain a satisfactory model with a finite amount of measured data, and the performance is better than that of existing algorithms.