Data Prediction with Few Observations Based on Optimized Least Squares Support Vector Machines

Abstract

In traditional statistics methods, large samples are needed for accurate function estimation and data prediction. Least squares support vector machines (LS-SVM's) is a new machine learning method for function estimation even with small samples. However, inappropriate LS-SVM's algorithmic parameters always bring poor results. In this paper, first a LS-SVM's algorithmic parameters optimization method is suggested which is called multi-layer adaptive best-fitting parameters search algorithm. Learning error of samples can be controlled to minimum by the new method. And then, a data prediction model based on the parameter-optimized LS-SVM's is approached,and the Ti-26 alloy material performance prediction experiments are analyzed with this model. The results show that the model has excellent learning ability and generalization and can provide more accurate data prediction only with fewer observed samples, as compared with supervised linear feature mapping (SLFM) neural network.