Research On Fuzzy Support Vector Machine Based On The Judgement Of Dual Membership

Support Vector Machine (SVM) is an algorithm which is developed in recent years and mainly used to solve classification problem. Due to the unique advantages, it has become a new research focus in the area of machine learning. However, there are still several flaws in SVM, in which the problem that SVM is very sensitive to noise and outliers in the training set is one of the primary reasons hindering its large-scale application.In order to solve this problem, a new fuzzy support vector machine algorithm based on the idea of affinity is proposed in this paper. Two concepts - "affinity" and "optimal radius" - are defined to assign the different membership functions due to the place where the sample is, inside or outside the optimal radius, which can make a good distinction between the effective samples and the noised ones. In the meantime, this paper also sets the formulas of two-way memberships, which evaluate two memberships, positive and negative, to each training sample in order to measure the influences of the same and opposite samples at the same time. Finally, on the basis of the foregoing preparation, two algorithms of determining fuzzy memberships and corresponding models are put forward respectively.The simulation experiments in MATLAB show the effectiveness and correctness of the methods:better ability to resist noises than traditional support vector machine and fuzzy support vector machine with the single membership.The proposed algorithm is based on the assumption:the classification of a sample is concerned with the Euclidean distance in the feature space where the samples are located. Hence, it is important that whether we can find the appropriate kernel function and make the samples to satisfy this assumption in the feature space. However, kernel function itself is a worthy research field, so our next step is to integrate the latest kernel theory achievement into the proposed algorithm this paper presents.