基于属性约减的助推技术及其应用

Attribute reduction-based boosting and its application

助推技术是机器学习和数据挖掘领域一种重要的方法,它能够大大提升预测精度,但往往容易造成训练过度,即训练精度过高导致模型外推性变差。本文提出了使用神经元网络技术进行属性约减后进行助推决策树建模的方法,较大程度上避免了助推的过度训练问题。在湖南某市电信数据库中进行了客户流失分析建模实验,结果表明该方法在模型的精度、结果的可理解性以及模型外推精度方面均优于同类算法。

Boosting is one of the most important methods in the field of data mining, which can improve the prediction accuracy of the algorithm. However, it often results in overfitting, namely, the deterioration of model＇ s extrapolation due to high accuracy. One of the causes of the overfitting in boosting is the static voting strategy. In this paper, an application of neural net on attribute reduction is proposed, by which, to a greater extent, the overfitting in boosting is avoided. We made an analysis of client lost based on the telecom database of a city in Hunan Province, which shows that this application effectively improves the accuracy of models and the comprehensibility of results.