题    目：Support Vector Machine with a Novel Slide Loss

主讲人：王宜举 教授

单    位：曲阜师范大学

时    间：2025年12月8日 15:00

地    点：九章学堂南楼C座302

摘    要：Among all the penalty functions for misclassified data points in typical support vector machine (SVM) models, the 0-1 loss function is an ideal one. However, its discrete nature presents computational challenges for the corresponding optimization model. To tackle this, various surrogates for the 0-1 loss functions are proposed, yet most involve a trade-off between Bayes optimality and computational efficiency. Different from these, in this talk, we introduce a novel slide loss function which is smooth and serves as an asymptotic approximation to the 0-1 loss function. Further, minimizing this loss function can yield a Bayes classifier associated with the 0-1 loss. Based on this loss, we develop a new type of SVM model which is a smooth unconstrained optimization problem. By investigating the property of the slide loss under the optimal condition of the model, we first derive a working set strategy and subsequently establish a numerical solution method for the model with reduced computational complexity. The numerical experiments made on simulated data and real-world data show validity of the proposed model and the efficiency of the proposed algorithm.

简    介：王宜举，曲阜师范大学教授，博士生导师，管理学院院长。中国科学院博士,香港理工大学和南京师范大学博士后。主要从事最优化问题的理论与算法研究，学术成果发表在SIAM J Matrix Anal Appl, Math Programming, IEEE Trans Signal Processing, J Graph Theory等学术期刊。主持国家自然科学基金5项,获教育部和山东省自然科学二等奖4项。多次到香港城市大学、香港理工大学和澳大利亚的科廷大学进行学术访问和交流。出版运筹学专业研究生教材《非线性最优化理论与方法》一部，先后被重庆大学、国防科技大学、北京交通大学、东北师范大学、海南大学、重庆师范大学、山东科技大学等高校聘为主讲教师为相关专业的研究生和博士生讲授《最优化理论与方法》。2015年享受国务院特殊津贴。