Fast and De-noise Instance Selection Method for SVMs Training Based on Clustering and Intuitionistic Fuzzy Number

Support vector machines (SVMs) are potent machine learning classifiers widely used in many real-world domains due to their solid theoretical foundation and excellent generalization performance. Instance selection is an effective method for SVMs to overcome the challenges of large datasets, in which the training efficiency of SVMs will decrease with the increase in data volume. However, the current instance selection methods are notably vulnerable to noise, often mistakenly identifying noises as ‘high-value data.’ In order to address the above shortcomings, a fast and de-noise instance selection method based on clustering and intuitionistic fuzzy number is proposed. Each instance is assigned with an intuitionistic fuzzy number, where the degree of membership hinges on the distance of each instance to an approximate hyperplane built based on cluster centroid, and the degree of nonmembership hinges on the number of dissimilar-class instances in the k-nearest neighbors of each instance. Then, the contribution of each instance is measured by the intuitionistic fuzzy number score function, which can more precisely characterize the distribution of support vectors and noises to preserve instances near the decision boundary and eliminate noises and redundant instances simultaneously. Experimental results with comparative methods on standard and noisy datasets indicate that the proposed method has good classification performance while significantly reducing training time and can effectively de-noise. Therefore, the proposed method can achieve higher generalization performance on large real-world datasets compared to other methods.