A Study Of The Prediction Modeling Based On Support Vector Machine

Prediction analysis is to anticipate and forecast the unknown in the future based on the known in the past and at present. It is a qualified and quantified description of undetermined and unknown incidence that is possible to happen during the development of the predicted incident. This paper presents an intelligence prediction technique -- an issue of classification and regression on the basis of support vector machine (SVM).SVM is the hot issue accompanying artificial neural network in machine learning. It involves any practical problems such as classification and regression estimation. SVM is based on the structural risk minimization (SRM) of the statistical learning theory (SLT), avoiding the overfitting and improving the generalization performance to some extent. The hybrid and ensemble method is recently one of the most popular fields in the machine learning. This paper mainly focuses on the prediction problem by the application of hybrid and ensemble thinking into the modeling base on SVM.The main work of this paper is as follows:We present a hybrid algorithm combining attribute reduction of RS with classification principle of SVM. Firstly, the attribute reduction of RS has been applied as preprocessor so that we can delete redundant attributes and conflicting objects from decision-making table from the 2-dimension point of view but efficient information lossless. Then, we implement classification modeling and forecast based on SVM. In the last, we give a computed example about the classification of inventory.Based on the theory above, we present a novel hybrid algorithm in case of numerous attributes. Firstly, we put forward two definitions about the selection of attribute importance: correlation degree and contribution degree. In the principal components analysis (PCA), we select the set of important attributes—the set of principal components based on the correlation degree that shows the interrelation among the attribute variances; and then we implement the attribute reduction again based on the contribution degree of selected attributes in the process of PCA to decision variances, deleting redundant and unimportant attributes and decreasing the dimension of the data inputted into SVM. Finally the simulation demonstrates the validity and accuracy of the method.We put forward a prediction modeling of constructing the SVM ensemble based on the thinking of ensemble learning. Assume that different individual SVM is trained independently using different training samples. Each individual SVM provides different generalization performance. The region of test examples resulting in lower error rate by each individual SVM can be very different each other and the region covered by SVM ensemble is extended further. This expansion implies that the SVM ensemble will improve the generalization ability further. In this paper, we propose to use different methods to aggregate the individual SVM. The methods include linear combination (majority voting and the least squares etc.) and nonlinear combination called as double-layer hierarchical combining that use another upper-layer SVM to combine several lower-layer SVMs. By constructing SVM ensemble, the operation speed is increased largely by freeing memory occupied with kernel function matrix during the quadratic programming solution and prediction accuracy is improved greatly. The various simulation results show that the proposed SVM ensemble is superior to single SVM.