【摘要】 代价敏感学习是机器学习中一个重要的领域.由Masnadi等提出的代价敏感的支持向量机通过将铰链损失函数代价敏感化来处理代价敏感问题,比传统的代价敏感学习方法具有更好的泛化精度.现实中的数据往往是通过在线增量式获取的,而传统的全量式学习算法每次增加样本时都需要重新从头计算,因此浪费了很多时间.为了使得代价敏感的支持向量机能够在在线学习的场景下具有更高的效率,提出了一种增量式的代价敏感支持向量机算法.该算法可以在新增样本时直接更新已有的训练过的模型,不需要从头开始重新训练.在多个数据集上的实验结果也显示出了该方法与传统的批处理方法相比,在速度上的具有显著的优势.更多还原

【Abstract】 Cost-sensitive learning is an important field in machine learning,which widely exists in realworld applications,such as cancer diagnosis,credit application,etc.Cost-sensitive support vector machine proposed by Masnadi et al.handles cost-sensitive problems through making the hinge loss function costsensitive,which has better generalization accuracy than other traditional cost-sensitive algorithms.In practice data are obtained one batch after another.Conventional batch algorithms would waste a lot of time when appending samples,because they should re-train the model from scratch.To make the cost-sensitive support vector machine more practical in on-line learning problems,an incremental cost-sensitive support vector machine algorithm was proposed,which can directly update the trained model without re-training it from scratch when appending samples.Experiment study on several datasets show that our algorithm is significantly more efficient than batch algorithms of the cost-sensitive support vector machine.更多还原