基于优化的xgboost模型的商业银行电话营销效果分析

发布时间：2018-05-08 15:43

本文选题：电话营销 + 数据挖掘　；参考：《兰州大学》2017年硕士论文

【摘要】：随着金融的全球化和自由化进程的加快,银行业的竞争越来越激烈,依靠存贷差的传统盈利模式已经很难再持续发展下去。在营销领域,传统的粗放式的客户营销策略转向精细化的客户营销策略,展开以客户为中心的精准营销活动已是大势所趋。所有银行营销活动都将依赖于其庞大的数据集,单纯地利用人工来分析这些数据是不可能的,数据挖掘模型有助于进行这些数据集的分析。本文以预测银行电话营销结果为目标,首先对研究问题的背景、意义、国内外研究现状以及研究方法与思路进行介绍。其次介绍了本研究中所涉及的数据挖掘技术,包括分类回归树、逻辑回归、随机森林、梯度迭代决策树等算法,在此基础上,介绍了xgboost集成学习框架。并介绍了处理不均衡数据集的边界合成少数类过抽样算法(BorderlineSMOTE),并将该算法与以上五种数据挖掘算法相结合,建立了银行电话营销分类模型。通过ROC曲线、AUC值、敏感度、特异度等指标发现,Borderline-SMOTE算法结合xgboost所得到的模型预测效果最佳,AUC值达到0.97。其次,xgboost模型不管是在预测效果还是运算效率上,都要优于本文构建的其它模型。本文还将两种信息提取方法(变量重要性分析和CART规则提取)用于提取数据集的关键信息,并揭示了几个关键属性(例如,Euribor3m、持续时间、年龄等)。这样的信息提取证实了所获得的模型对于电话营销活动管理者是可信的和有价值的。
[Abstract]:With the acceleration of financial globalization and liberalization, the competition of banking is becoming more and more fierce. It is difficult to continue to develop the traditional profit model which depends on the difference between deposit and loan. In the field of marketing, the traditional extensive customer marketing strategy turns to the refined customer marketing strategy, and it is the trend of the times to launch the client-centered precision marketing activities. All marketing activities of banks will depend on their huge data sets. It is impossible to analyze these data simply by using manpower. The data mining model is helpful for the analysis of these data sets. This paper aims at predicting the results of bank telephone marketing. Firstly, it introduces the background, significance, current research situation, research methods and ideas of the research. Secondly, this paper introduces the data mining techniques involved in this study, including classified regression tree, logical regression, stochastic forest, gradient iterative decision tree and so on. On this basis, the integrated learning framework of xgboost is introduced. This paper also introduces the borderline SMOTET algorithm of edge synthesis for dealing with unbalanced data sets, and establishes a bank telephone marketing classification model by combining this algorithm with the above five data mining algorithms. Through the ROC curve, sensitivity, specificity and other indicators, we found that Borderline-SMOTE algorithm combined with xgboost has the best prediction effect of 0.97. Secondly, the xgboost model is superior to the other models in terms of prediction effect and operational efficiency. In this paper, two information extraction methods (variable importance analysis and CART rule extraction) are used to extract the key information of the data set, and several key attributes (such as Euribor3m, duration, age, etc.) are revealed. Such information extraction confirms that the obtained model is credible and valuable to telephone marketing managers.
【学位授予单位】：兰州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：F274;F831.2

【参考文献】