The Research On Incorporating Method Of Prior Knowledge And SVM

Support vector machine (SVM) is one type of learning machines that is paid wide attention in recent years. Based on statistical learning theory (SLT), SVM possesses many merits such as concise mathematical form, standard fast training algorithm and excellent generalization performance, so it has been widely applied in data mining problems such as pattern recognition, function estimation and time series prediction, et al. However, some problems still need to be resolved in SVM researches, for example, the model selection, efficiency of SVM for large-scale training set, and incorporating prior knowledge into SVM etc. Generally, available information at hand about problem is obtained from training set in SVM. If prior knowledge can be incorporated into Support Vector learning machines, the generalization performance of SVM may be improved efficiently. Therefore, how to obtain the prior knowledge and incorporate them into SVM is an important issue in the domain of SVM researches. In the thesis, the problem of incorporating prior knowledge into SVM is investigated systematically. The main achievements are concluded in the following:(1) Analyzes the existing methods for incorporating prior knowledge into SVM.(2) Proposes a new incorporating approach, which incorporates invariance into a pattern recognition classifier through representing the trajectory manifold of invariance transformation by the best approximate point. Prior knowledge about invariance is very useful in classification problems, and it is an important focus for prior knowledge incorporating researches. Although there are some approaches reported presently, these approaches have respective disadvantages such as too strict constraints for invariant transformation, or great computation cost, or rather intractable implementation. In the thesis, the approach incorporating invariance into SVM is presented based on the best approximate point, which makes use of the maximum margin hyperplane principle of SVM. Hence, the presented approach can avoid the great computation cost and improve the generalization performance of SVM effectively.(3) Provides the incorporating method of invariance and SVM model based on the best approximate point. The presented incorporating approach is testified on a real-world dataset, and the experiment results demonstrate the effectiveness and practicability of the model.(4) Presents that the time correlation for time series data can be used as a kind of prior knowledge for time series prediction problem, and proposes a new approach of kernel function construction, namely time series kernel, which can incorporate time correlation into SVM, and obtain good generalization performance.(5) Introduces time series kernel into the domain of environmental time series modeling. The application results demonstrate that, compared with the traditional kernel function, time series kernel possesses more accurate predicting capability.The research work in the thesis is the one of key problems in SVM researches. The obtained results not only have important theoretical significance, but also possess direct application value for real-world problems.