摘要：

为使智能家居系统从传感器网络返回的数据中自动识别用户行为并生成个性化服务策略,提出一种引入惩罚项的随机森林算法。对每次迭代过程中使用的属性集设置不同的惩罚项因子,生成尽可能不同的决策树,从而兼顾集成算法的多样性与分类精度。在UCI、CASAS数据集上的实验结果表明,与传统集成分类算法Bagging、Adaboost相比,该算法具有更高的分类精度与噪声鲁棒性。

关键词: 自动识别, 集成学习, 惩罚项因子, 分类精度, 噪声鲁棒性

Abstract:

In order to enable smart home systems to automatically identify user behavior and generate personalized service strategies from data returned by sensor networks,a random forest algorithm with penalty items is proposed.Different penalty factors are set to the attribute set used in each iteration to generate as different decision trees as possible,keeping the diversity and classification accuracy of the ensemble algorithm.Experimental results on UCI and CASAS datasets show that compared with the traditional integrated classification algorithms,Bagging and Adaboost,this algorithm has higher classification accuracy and noise robustness.

Key words: automatic recognition, ensemble learning, penalty factor, classification accuracy, noise robustness

中图分类号: 

• TP391