Abstract: Time adaptive support vector machine and its improved method has shown preferable performance on data stream classification by simultaneously solving several sub-classifiers locally and globally in an alternative kernel space to enforce the sub-classifiers evolve stably. However, due to the randomness of data in many practical scenarios, concept drift in steady rhythm cannot be guaranteed, thereby resulting in that such instability of sub-classifier serials should be consider in TA-SVM method. In this paper, in order to loosen the constrains on normal vector and bias and margin of sub-classifiers, a novel classifier named Flexible Drift Support Vector Machines（FD-SVM） is proposed by handling normal vector and bias discriminatory when sub-classifier drifting. The effectiveness of the proposed FD-SVM is experimentally verified.

Key words: data stream classification, sub-classifiers serials, support vector machine, flexible drift

摘要： 针对数据流分类，时间自适应支持向量机（Time Adaptive Support Vector Machine，TA-SVM）及其改进方法通过在核空间上协同求解多个子分类器而取得了较好的性能，其原理是在局部优化的同时兼顾全局优化，强制子分类器序列稳定地变化。然而在很多应用场景中，由于数据具有一定的随机性，难以确保概念模型以不变的节奏漂移，因而TA-SVM分类模型中应充分考虑子分类器序列的不稳定性。为了放松TA-SVM方法对子分类器序列的约束，使子分类器法向量、偏置量的变化具有更大的灵活性，提出了柔性漂移支持向量机（Flexible Drift Support Vector Machines，FD-SVM），在继承TA-SVM方法协同求解思想的基础上，灵活对待子分类器变化。实验结果表明，FD-SVM方法能有效提升对非静态数据的分类性能。

关键词: 数据流分类, 子分类器序列, 支持向量机, 柔性漂移