The Research On Data Fusion Strategies Based On Support Vector Machine

Machine learning algorithms have got extensive attentions of scholars and have been applied in various fields in recent years. At the same time, data fusion methods have been used in a great deal of practical application in the past decade. If these two methods are combined together to make up each other's defects and keep their advantages, then the intelligent methods can be further improved.Support vector machine is one of the most popular machine learning algorithms. Support vector machine is derived from Vapnik's statistical learning theory, and the core of it is Vapnik Chervonenkis dimension and structural risk minimization theory. It has good generalization ability, applied to small samples, and is able to avoid local optimal solution, so it can be used well to resolve problems of classification and regression. The most important thing in implementing support vector machine is its training algorithm. There are some training algorithms now, but the most popular one is the sequential minimal optimization algorithm. Support vector machine is developed for binary classification at first. But in real world, we usually have to classify samples belong to more than two classes, so scholars propose some methods to construct multi-class support vector machine. And there are some strategies to make support vector machine outputs probabilistic results.Data fusion strategy has widely applied in the military and civilian areas in recent years, and there are a lot of data fusion methods. The most commonly used data fusion methods are Bayesian theory and Dempster-Shafer theory. The basis of Bayesian theory is probability theory. And the Dempster-Shafer theory Dempster-Shafer theory is applied to uncertain cases, it can be seen as the extension of Bayesian theory. So these two methods are both used widely.The purpose of this paper is to combine support vector machine and other data fusion methods, then both of these two technologies can be completed to obtain a better classification methods. In this paper, some multiple sources data fusion strategies base on MSVM are proposed, including sum strategies, Bayesian strategies, DS strategies and 2-Layer MSVM strategies. After the analysis of these theoretical feasibility about these data fusion strategies, they are implemented by using libSVM. Then many experiments are taken placed. From the experiments' results, we can see that the data fusion strategies proposed in this paper can get better classification results. These data fusion strategies have a distinct advantage—they can handle small sample classification problems, especially with the DS theory. Support vector machine is fit for small data sets, and DS theory shows good performance about uncertain cases, so the experimental results proved the correctness of these two theories. These can further explain that the data fusion strategies based on support vector machine proposed in this paper are fit for classification with small sample.