Abstract：As a popular machine learning algorithm, the standard support vector machine(SVM) is faced with two problems: 1) how to effectively use the local information of data manifold; 2) the classification hyperplane sensitive to the outliers in the data. Therefore, a learning algorithm called adaptive local graph embedding weighted-penalty support vector machine(ALGEWP-SVM) is proposed. On the condition of keeping the optimization framework of the standard SVM,
the proposed algorithm joins the requirements of maximizing inter-class margin of the entire data and optimizing local distribution of the data manifold in the objective function, which optimizes the hyperplane of classification decision and simplifies the process of kernelization. Meanwhile, the proposed algorithm introduces the global structure information of data to automatically repress the influence of the outliers upon the hyperplane and improve the robustness of the algorithm. The results of the experiment on artificial, standard and image datasets show the effectiveness of the proposed algorithm.