关键词: 支持向量数据描述, 增量学习, 自适应阈值, 淘汰机制

Abstract: In incremental learning based on Support Vector Data Description（SVDD）, existing methods often discard more useful samples for new models which lead to serious drop in the accuracy of incremental learning. Focused on this problem, a new removing algorithm for incremental SVDD learning is proposed. Based on the analysis of support vector set’s change in incremental learning process, adaptive learning threshold [α] is defined to filter out samples that may become new support vectors. Furthermore, a sample removing method is proposed to avoid repetitive training for useless samples. Simulation with the UCI benchmark datasets and warhead target simulation HRRP data indicates that new algorithm has better classification accuracy than conventional incremental algorithm while improving the training efficiency of the model.

Key words: Support Vector Data Description（SVDD）, incremental learning, adaptive threshold, removing method