基于GBDT的商品分配层次化预测模型  

GBDT based hierarchical model for commodity distribution prediction

作　　者：朱振峰[1,2] 汤静远 常冬霞 赵耀[1,2] ZHU Zhenfeng;TANG Jingyuan;CHANG Dongxia;ZHAO Yao(School of Computer and In fo rmation T -c chn ology;Bcijing Key Laboratory of Advanced In fo rma tio n Science and Ne two rk Tcchnolo gy,Beijing Jiaotong University,Beijing 10004 4,China)

机构地区：[1]北京交通大学计算机与信息技术学院,北京 100044 [2]北京市现代信息科学与网络技术重点实验室,北京100044

出　　处：《北京交通大学学报》2018年第2期9-13,45,共6页JOURNAL OF BEIJING JIAOTONG UNIVERSITY

基　　金：国家自然科学基金(61572068,61532005);教育部新世纪优秀人才支持计划项目(NCET-13-0661);中央高校基本科研业务费专项资金(2015JBM039)

摘　　要：商品预测是使用以往商品信息去估计和推断未来商品的销售趋势,并以此作为对商品进行合理调配与规划的依据.为实现对商品销售的精确预测,在GBDT基础上,提出了一种层次化集成预测模型(HGBDT).针对数据表征的高维问题,基于Bagging思想,在特征空间构建了该模型,实现对商品的有效描述,以此提高预测模型的性能与泛化能力.在开放数据库上的实验结果验证了本文模型的有效性.Commodity prediction uses the previous commodity information to estimate and infer the future trends of the commodity,and it can be used for carrying out reasonable planning and distribution of commodity. To achieve accurate forecast of merchandise sales？ a commodity distri-bution prediction model （ HGBDT ） based on Gradient Boosting Decision Tree （G B D T ） is proposed. To alleviate the problem of dimensionality curse,w e construct a Bagging based hiera-- chical ensemble learning model.The temporal-spatial property of commodity is exploited for char-acterizing commodity effectively, which is beneficial to boost the generalization of the learned prediction model. Experimental results on open dataset demonstrate the effectiveness of the pro-posed method.

关 键 词：决策树 回归模型 GBDT 集成学习 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]