Papers by Author: Kai Li

Abstract: To improve the generalization performance for ensemble learning, a subgraph based selective classifier ensemble algorithm is presented. Firstly, a set of classifiers are generated by bootstrap sampling technique and support vector machine learning algorithm. And a complete undirected graph is constructed whose vertex is classifier and weight of edge between a pair of classifiers is diversity values. Secondly, by searching technique to find an edge with minimum weight and to calculate similarity values about two vertexes which is related to the edge, vertex with smaller similarity value is removed. According to this method, a subgraph is obtained. Finally, we choose vertexes of subgraph, i.e. classifiers, as ensemble members. Experiments show that presented method outperforms the traditional ensemble learning methods in classification accuracy.

Abstract: By combining fuzzy support vector machine with rough set, we propose a rough margin based fuzzy support vector machine (RFSVM). It inherits the characteristic of the FSVM method and considers position of training samples of the rough margin in order to reduce overfitting due to noises or outliers. The new proposed algorithm finds the optimal separating hyperplane that maximizes the rough margin containing lower margin and upper margin. Meanwhile, the points lied on the lower margin have larger penalty than these in the boundary of the rough margin. Experiments on several benchmark datasets show that the RFSVM algorithm is effective and feasible compared with the existing support vector machines.