基于高维精准画像与深度学习的卷烟零售户动态分档研究  

Research on dynamic classification of cigarette retailers based on high-dimensional precise profiling and deep learning

作　　者：谭琛[1] 周欣然 栾晓宇[1] 李敏刚[2] 杨云丽 顾敏洁 金雅昭 TAN Chen;ZHOU Xinran;LUAN Xiaoyu;LI Mingang;YANG Yuni;GU Minjie;JIN Yazhao(Information Center,Shanghai Tobacco Group Co.,Ltd,Shanghai 200082,China;Marketing Center,Shanghai Tobacco Group Co.,Ltd,Shanghai 200082,China;Shanghai Pintechs Technology Co.,Ltd,Shanghai 201101,China;Research Center of Digital Transformation for Enterprises,Beijing Institute of Big Data Research,Beijing 100080,China)

机构地区：[1]上海烟草集团有限责任公司信息中心,上海市长阳路717号200082 [2]上海烟草集团有限责任公司营销中心,上海市长阳路733号200082 [3]上海品见智能科技有限公司,上海市合川路2679号201101 [4]北京大数据研究院企业数智化转型研究中心,北京市海淀区颐和园路5号100080

出　　处：《中国烟草学报》2024年第3期116-124,共9页Acta Tabacaria Sinica

摘　　要：【目的】丰富卷烟零售户考核评价的维度,建立档位动态认定的科学标准,辅助档位调整、分类管理及投放策略制定等营销工作。【方法】提出基于高维精准画像体系和深度学习模型进行档位识别的解决方案,介绍档位标签推演、画像特征指标设计、识别模型搭建、新零售户样本预测及评价、档位动态调整等环节的可行策略,并通过分析特征重要性挖掘到零售户档位的潜在影响因素。【结果】在数值实验部分,基于A、B市场多个投放周期内的营销数据,对比非线性深度学习DXGBOOST模型与线性回归LR模型、梯度提升XGBOOST模型的分类识别效果:相比线性的LR,DXGBOOST算法可将A、B两市零售户档位识别准确率分别提升9.09%和15.43%;相比非线性的XGBOOST,可进一步提升精准档位及档位区间的准确率,降低误判的偏差程度。【结论】实践证明,在保证稳定性的前提下,基于DXGBOOST模型的档位识别方法可以很大程度提升卷烟零售户档位的划分精度,为及时开展科学、精准、智能地营销分析与商业决策提供数据和理论支撑。[Objective]This study aims to enrich the evaluation dimensions of cigarette retailers,establish the scientific standards for dynamic identification,and assist the marketing works such as grade adjustment,classification management,and delivery strategy formulation.[Methods]this paper introduces the overall solution for cigarette retailers identification,which is based on a high-dimensional precise profiling system and the deep learning model.Some effective strategies have been put forward,such as deducting the grade for retailers,designing the indicators of portrait,building the identification models,predicting and evaluating new samples,adjusting grade dynamically.In addition,by analyzing the importance of features,the potential influencing factors of retailers are excavated.[Results]In the numerical experiments,based on marketing data from multiple launch cycles in Markets A and B,a comparison of the classification recognition effects of the nonlinear deep learning DXGBOOST model with the linear regression LR model and gradient boosting XGBOOST model is presented.Compared with the linear LR,the DXGBOOST algorithm can increase the grade identification accuracy of retailers in Markets A and B by 9.09%and 15.43%respectively.Compared to the nonlinear XGBOOST,it further enhances the accuracy of precise grades and grade intervals,reducing the degree of misjudgment bias.[Conclusion]Practice has proven that,while ensuring stability,the grade identification method based on the DXGBOOST model can greatly improve the grading accuracy of cigarette retailers,providing data and theoretical support for timely conducting scientific,precise,and intelligent marketing analysis and business decision-making.

关 键 词：精准画像 档位识别 深度学习 DXGBOOST模型 卷烟零售户 

分 类 号：F721[经济管理—产业经济] TP18[自动化与计算机技术—控制理论与控制工程]