The Theory And Applications Of Autoregulative Fuzzy Decision-making SVM

Support Vector Machine (SVM) is a powerful learning machine with a universalfeed-forward network structure. Recently, SVM has been widely used in many actualapplications, especially for dealing with many classification problems. In this thesis, ourresearch also focuses on the real-world data classification problems. Two hidden influencesmainly exist in the real-world data.Firstly, the existence of interaction and noises in the real world usually disturbs theoriginal data, especially in the zone around the separating boundary between different subsets.Ordinarily, as a result of these disturbances from noises, this zone may become from a clearseparation zone to a gray one, which increases the difficulty for classification of the databaseand the performances of the classifiers also become worse apparently.Secondly, in the actual applications, the unbalance problem is a common phenomenon.That means that one of the classes in the binary-labeled real-world data sets is usually muchlarger than the other one. The reason of which is that in most real-world problems, thefrequency of a event is much lager than the opposite case mostly. This imbalance situationalways causes the excursion of the boundary in classification problems.In order to reduce the influences caused by interaction and noises existed in thereal-world datasets and imbalance between different classes, we proposes an improved model.Being different from traditional SVM classifiers, the new model takes the thought about fuzzytheory into account. A fuzzy decision-making function is built to replace the sign function inthe prediction stage of classification process. In the prediction part, the decision values areused to construct the fuzzy decision-making function. In addition, a boundary offset is alsointroduced to modify the boundary excursion. By calculating the weighted harmonic mean ofall decision values of support vectors, an accurate offset value can be gotten. On account ofthis offset, the boundary is set to an optimal position.This flexible design of this autoregulative fuzzy decision-making SVM model candescribe the properties of real-world conditions more correctly. Some better and more robustperformances are presented in simulations.