基于优化的xgboost-LMT模型的供应商信用评价研究

发布时间：2018-04-15 12:06

本文选题：供应商信用评价 + xgboost　；参考：《广东工业大学》2016年硕士论文

【摘要】：随着全球经济迅猛发展,市场竞争已经发展成为供应链与供应链之间的竞争。供应链管理模式作为一种新的管理理念与模式,近年来在国内外逐渐受到重视,被视为是企业提升市场竞争力的助推剂。在供应链管理模式中,拥有优秀的供应商合作伙伴是企业竞争致胜的关键,而对供应商信用进行正确评价是企业组建优秀供应商团队的重要因素。因此,对供应商的信用进行准确评价具有重要意义。基于供应商信用评价在大数据环境下的实质为一种模式识别问题,分类机的性能表现在供应商的信用评价问题中至关重要。xgboost集成学习框架是2015年被提出种新的集成学习模型,它以其出色分类性能和鲁棒性在模式识别问题上得到了较为广泛的使用。鉴于此,本文尝试将xgboost集成学习框架应用于解决供应商的信用评价问题。xgboost框架在使用时,需要设置模型参数收缩步长和子节点中最小的样本权重阀值。通常,参数组合的选取对模型的分类性能影响较大。以往的参数寻优方法,主要依据经验给定或实验穷举,但带有一定的不确定性和随机性。本文将人工化学反应优化算法(Artificial Chemical Reaction Optimization Algorithm, ACROA)应用于xgboost框架的参数寻优,并设定不同的初始反应物个数进行数值实验,验证基于ACROA算法的参数寻优的稳定性。LMT (Logistic Model Tree)算法是决策树的一种扩展算法,它将普通的决策树模型和Logistic回归模型相结合,比单独的决策树模型和Logistic回归模型具有更强的分类性能。为进一步提高LMT算法的分类性能和解释能力,本文将反正切Lasso惩罚函数引入LMT模型,并通过实验验证反正切Lasso对LMT算法分类性能的优越性。为提高供应商信用评价中分类模型的分类精度和鲁棒性,本文提出了一种基于优化的xgboost-LMT模型——使用ACROA算法对xgboost框架进行参数寻优,并将基于反正切Lasso惩罚函数的LMT算法引入xgboost框架,作为框架的基分类器(Based Learner)。本文在两个真实的供应商信用数据集上进行对比实验,对比算法包括本文提出的基于优化的xgboost-LMT模型、独立基于反正切Lasso惩罚函数的LMT模型、支撑向量机、人工神经网络和C4.5决策树。实验结果表明：本文提出的供应商信用评价算法有效可行,在准确度、查全率和AUC方面,比其它四种算法有明显提升；并且,当数据集中包含异常值时,本文提出的分类算法具有良好的鲁棒性。
[Abstract]:With the rapid development of global economy, market competition has become the competition between supply chain and supply chain.As a new management concept and model, supply chain management has been paid more and more attention at home and abroad in recent years, and is regarded as a booster for enterprises to enhance their market competitiveness.In the mode of supply chain management, having excellent supplier partners is the key to the success of enterprise competition, and the correct evaluation of supplier credit is an important factor for enterprises to set up excellent supplier team.Therefore, it is of great significance to evaluate the credit of suppliers accurately.Based on the fact that supplier credit evaluation is a kind of pattern recognition problem under big data environment, the performance of classifier is very important in supplier credit evaluation. Xgboost integrated learning framework is a new integrated learning model proposed in 2015.It is widely used in pattern recognition for its excellent classification performance and robustness.In view of this, this paper attempts to apply the xgboost integrated learning framework to solve the supplier credit evaluation problem. When the xg boost framework is used, it is necessary to set the model parameter shrinkage step size and the minimum sample weight threshold in the sub-nodes.Usually, the selection of parameter combination has a great influence on the classification performance of the model.Previous parameter optimization methods are mainly based on empirical or experimental exhaustive, but with certain uncertainty and randomness.In this paper, artificial Chemical Reaction Optimization algorithm (ACROAA) is applied to the parameter optimization of xgboost framework, and different initial reactants are set up for numerical experiments.It is proved that the stability of parameter optimization based on ACROA algorithm. LMT / Logistic Model tree algorithm is an extended algorithm of decision tree. It combines the ordinary decision tree model with the Logistic regression model.It has better classification performance than the single decision tree model and Logistic regression model.In order to further improve the classification performance and interpretation ability of the LMT algorithm, this paper introduces the Lasso penalty function into the LMT model, and verifies the superiority of the Lasso in the classification performance of the LMT algorithm by experiments.In order to improve the classification accuracy and robustness of classification models in supplier credit evaluation, this paper proposes a xgboost-LMT model based on optimization, which uses ACROA algorithm to optimize the parameters of xgboost framework.The LMT algorithm based on the Lasso penalty function is introduced into the xgboost framework as the base classifier of the framework.This paper carries on the contrast experiment on two real supplier credit data sets. The contrast algorithm includes the optimized xgboost-LMT model, the independent LMT model based on the Lasso penalty function and the support vector machine.Artificial neural network and C4.5 decision tree.The experimental results show that the proposed supplier credit evaluation algorithm is effective and feasible, and its accuracy, recall and AUC are significantly higher than those of the other four algorithms, and when the data set contains outliers,The classification algorithm presented in this paper has good robustness.
【学位授予单位】：广东工业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：F274;TP311.13

【相似文献】