基于神经网络模型的纳税评估及实证研究  

Tax payment assessment and an empirical study based on neural networks model

作　　者：陈凯明[1] 邬春学[1] 高丽萍[1] 

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《信息技术》2015年第11期49-54,共6页Information Technology

基　　金：国家自然科学基金项目(61202376);上海市教委科研创新项目(13YZ075)

摘　　要：提出一种基于神经网络模型的企业纳税状态评估的操作方法,并以上海市某区木制家具制造企业纳税状态为例进行实证研究。首先,对数据进行筛选并对用于建模的数据采用自组织特征映射网络进行聚类,将其合理分为三组样本:训练样本、检验样本和测试样本。其次,同时使用训练样本和检验样本数据建立神经网络模型,并用测试样本对模型性能进行分析。最后,同时建立三种传统统计模型作为对比。分析发现,采用文中的方法建立的神经网络模型是可靠的,对企业纳税状态的识别更加准确,尤其是对非诚实纳税企业的识别风险度更低,神经网络模型具有很好的泛化能力。This paper puts forward an operation method of tax assessment on the enterprise, and conducted an empirical analysis based on the data of wooden furniture manufacturing enterprise＇ s tax state located in some district of Shanghai City. First of all, it selected data and used the self-organizing feature map network to cluster modeling data, then divided them into three groups of similar samples, training sample, validation sample and test sample. Secondly, it used training sample and validation sample together to train neural network model, used test sample to analysis model＇ s performance. At last, it also established three traditional statistical models for comparison. Through the analysis it found that the neural network model is more reliable, identify tax state of enterprises is more accurate, especially for non-honest enterprise identification. Thus the neural network＇ s risk of identify the tax state of enterprises is lower, and the neural network model has good generalization ability.

关 键 词：纳税评估 样本合理分组 神经网络模型 传统统计模型 实证研究 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]