考虑错分代价的LogitBoost算法及其在手机用户价值分类中的应用

针对传统LogitBoost算法将正确分类与错误分类平等看待且损失函数不收敛于代价敏感的贝叶斯决策问题.本研究在传统LogitBoost算法的基础上给出了一种基于错分代价改进的LogitBoost算法,并以某移动通讯公司的手机用户数据为基础检验了该算法的有效性,研究发现:与其他同类算法相比考虑错分代价后的LogitBoost算法的分类效果提升较明显;随着错分代价比值的增大,预期风险(同一阈值下)增大;同一错分代价比值下的预期风险,随着错分代价比值的增大表现出先增大后减小的趋势.该结论不仅说明了通过引入错分代价能有效降低模型的预期风险(这使新算法由关注分类错误率最小化转向关注预期风险最小),也为通讯公司提供了分析框架和决策参考.

LogitBoost algorithm considering the cost of misclassification and its application in the classification of mobile user value

In the traditional LogitBoost algorithm, correct classification and error classification are treated equally, and the loss function does not converge to the cost-sensitive Bayesian decision. Based on the traditional LogitBoost algorithm, we propose a LogitBoost algorithm that takes the misclassification costs penalty into account. We verify the validity of the algorithm by using mobile phone user data of a mobile communication company. The results show that:Compared with the other similar algorithms, the classification improvement effect of the LogitBoost algorithm considering misclassification costs is obvious. As the misclassification cost ratio increases, the expected risks (under the same threshold) increase. The expected risks under the same misclassification cost ratio have a tendency of increasing first and then decreasing with the increase of the misclassification cost ratio. This conclusion not only shows that the introduction of misclassification costs can effectively reduce the expected risk of the model (this makes the new algorithm shift the focus from minimizing the misclassification cost ratio to minimizing the expected risks), but also provides an analysis framework and decision reference for communications companies.