The Classification Method Study of Two-Dimension Iteration Transductive Support Vector Machine

Because of the restriction caused by the practical application, the semi-supervised learning is generally adopted in pattern recognition. The paper proposed the method, which takes the Support Vector Machine (SVM) as a basic tool with absorbing the transductive inference to make the full use of concerned training sample information so that high-quality separating hyperplane of transductive sample set can be generated. The method is composed of two transductive inferences that 1) based on time-varied separating hyperplane the number of labeled transductive samples is increased while the samples in the training set is simultaneously deleted, so-called y-dimension iteration transduction and 2) labeled training sample subsets are offered the gradually increasing punishment factor coefficients as well as the punishment factor of training set is decreased by steps, the so-called x-dimension iteration transduction. During the period x-dimension iteration transductions are nested in y-dimension iteration transductions, which end at the time when the training set is empty. Simulation proves the method has the higher accuracy and stronger robust than correlation knowledge induction SVM and prior knowledge induction SVM in the situation where the label of training set is unavailable, such as remote sensing thermal imaging cognition.