增量式Huber-支持向量回归机算法研究

A Study of Incremental Huber-Support Vector Regression Algorithm

传统的面向支持向量回归的一次性建模算法中样本增加时,均需从头开始学习,而增量式算法可以充分利用上一阶段的学习成果。SVR的增量算法通常基于ε-不敏感损失函数,该损失函数对大的异常值比较敏感,而Huber损失函数对异常值敏感度低。所以在有噪声的情况下,Huber损失函数是比ε-不敏感损失函数更好的选择,在现实情况当中。基于此,本文提出了一种基于Huber损失函数的增量式Huber-SVR算法,该算法能够持续地将新样本信息集成到已经构建好的模型中,而不是重新建模。与增量式ε-SVR算法和增量式RBF算法相比,在对真实数据进行预测建模时,增量式Huber-SVR算法具有更高的预测精度。

In the traditional support vector regression-oriented one-time modeling algorithm,when the number of samples increases,it is necessary to start from scratch,while the incremental algorithm can make full use of the learning results of the previous stage.The incremental algorithm of SVR is usually based on theε-insensitive loss function,which is more sensitive to large outliers,while the Huber loss function is less sensitive to outliers.So in noisy situations,the Huber loss function is a better choice in real-world situations.Based on this,this paper proposes an incremental Huber-SVR algorithm,which can continuously integrate new sample information into the already constructed model instead of remodeling.Compared with the incrementalε-SVR algorithm and the incremental RBF algorithm,the incremental Huber-SVR algorithm has higher prediction accuracy when performing predictive modeling on real data.